# 367: Claude introduces DLP, I thought it always stole Data
Duration: 46 minutes
Speakers: Matt, Justin, Matt Kohn, Ryan Lucas
Date: 2026-08-20

## Transcript

[00:07] Matt: Welcome to The Cloud Pod, where the forecast is always cloudy. We talk weekly about all things AWS, GCP, and Azure.

[00:14] Justin: We are your hosts, Justin, Jonathan, Ryan, and Matt.

[00:18] Matt: Episode 367, recorded for August 11th, 2026. Claude introduces DLP. I thought it always stole data though. Good evening, Ryan and Matt. How are you doing? Doing good.

[00:32] Matt Kohn: Doing well. Doing well.

[00:34] Matt: I did see that our show note writer Heather did call you out for your Eagles fandom again, which I do enjoy every time she does it.

[00:43] Ryan Lucas: Yeah.

[00:45] Justin: Problem is, is I'm just beaten at this point. Like I have no words. I'm deeply offended and just cut to the bone, and I just don't know how to respond. I just—

[00:56] Matt Kohn: I thought you just, you know, took the ownership of it at this point and became an Eagles fan because of it.

[01:02] Justin: Uh, I will never become an Eagles fan, but it does make me want to form like a— what's the opposite of a tribute band?

[01:09] Matt: Like a hate band?

[01:11] Justin: Yeah, which would be pretty funny. An Eagles hate band would be pretty funny. So like, maybe I'll use this as inspiration to, uh, to, you know, go write my own versions of their terrible music. I mean, you can use Suno to help you out with that, so I could, I could.

[01:27] Matt: Yeah, yeah. So there's lots of ways to make AI music these days to, uh, fulfill this fantasy. So, all right, well, we have a bunch of news to get to this week, and we should probably jump into it before we all turn into pumpkins at midnight around here. First up in AI is how machine learning makes money, uh, inference hooks. Are finally here. Anthropic is launching inference hooks in beta for Claude Enterprise, providing inline data loss prevention across chat, Claude Code, Claude Cowork, and other enterprise services through a single configuration point. The technical approach, uh, every inference request routes through a signed WebSocket connection to a customer-controlled security server, and Claude sends the prompt and context before the generation begins and waits for an allow/deny verdict before proceeding. Same section applies to tool call responses, including those from MCP connections. Skills, and plugins. The feature uses an open webhook-based protocol with a published schema, allowing integrations with existing DLP vendors such as Netscope, Palo Alto Networks, Proofpoint, and Zscaler, or a custom in-house security server without requiring separate per-product integration work. The rollout controls include shadow mode, log without blocking, role-based exclusions, and percentage-based rollouts, along with configurable failure policy, tolerance, and timeouts to match organizational risk requirements. Uh, this addresses a gap where inline enforcement was previously limited to Cloud Pod's client-side hooks, giving compliance teams a unified enforcement layer for sensitive data. Across cloud enterprise.

[02:48] Justin: Yeah, trying to configure these, you know, on the client side seemed impossible if you were worried about, you know, proprietary data being leaked out or, you know, people, you know, accidentally sharing information that they shouldn't. So I do think that this is, you know, a good tool to have like at an enterprise level. I am very curious to see what it does to response time I can see how this might add latency in between the, you know, the asker and the response. But I do think, you know, using webhooks is the, you know, the best way to keep that down. So, but then also it's subject on, you know, whatever you're having doing the analysis of the prompt. So let's see.

[03:31] Matt: I mean, AI is not known for its rapid real-time speed anyways. So, you know, if you added a few seconds for it to do this and like if you're using our next feature we'll talk about here in a second, auto mode. This is kind of the same mechanism that auto mode is in place for Anthropic Claude where it'll basically check the command and see if something that's in its known risky or is this a risky type query. And so it's similar kind of idea, just using it for DLP. Yep.

[03:57] Justin: I mean, you can do any kind of evaluation 'cause it's just going through the data. It is just the whole conversation in payload construct. So DLP is one aspect of it. You can anything, you could evaluate the speech or the, the sentiment or so.

[04:16] Matt Kohn: If given that they're kind of integrating all these third-party DLP tools into it also, it's going to be fun when you have your first outage because your Palo Alto isn't responding back and now Claude's down and people are yelling at you, not that your internet's down or anything else, but Claude is down and it's because your other third party. So it's going to be a fun debugging world at times when like random subsystems fail. And that's what I kind of worry about sometimes when you start daisy chaining all these tools together. You know, while I understand the need for it, you know, it's slowly going to be fun to kind of figure all these things out, especially some of these aren't the most modern DLP vendors that they have here. They've kind of added on these products over time. And as you guys said, speed and everything else start kind of You know, security is going to say day one, hey, we need to do all these things and without setting it up and validating and testing it. Looking at you, Ryan.

[05:11] Justin: Well, I mean, Anthropic has their own issues, so you could just blame them every time if it had slowness.

[05:15] Matt Kohn: I mean, they're perpetually down.

[05:18] Matt: I mean, I, I mean, I think the reality is, uh, unless you're using like a pretty antiquated DLP like Symantec, for example, I mean, most of the new ones like CrowdStrike's DLP, Palo Alto's, Netscope's, etc., they're all mostly inline as well in the agent already on the laptop. So if you could plug it into that in some way, or, you know, at least have it be configurable through that method, I think there's ways to make it seem less slow than it is.

[05:40] Justin: Yeah.

[05:41] Matt: I mean, and then a lot of companies are just pushing all of their AI traffic through proxies as well. So they're already adding latency. And that's also a point where you probably put some type of DLP injection as well. Yeah.

[05:51] Justin: I mean, yeah, you see Netscope and Zscaler as the launch partners on this. So it's. Makes sense.

[05:56] Matt: Yep. Makes total sense. Well, starting August 14th, Claude code will default to auto mode for Pro, Max, and Team plans, replacing manual permission prompts with a classifier that evaluates each tool call for irreversible, destructive, or external-facing actions. Enterprise API, Bedrock, and other cloud hardware integrations remain opt-in for now, with default rollout planned in the coming month. Anthropic's testing found manual review is less reliable than the classifier, with paid testers caught only 13.6% of injected dangerous commands, while auto mode blocked 89% of the same set. Human approval rates also declined as sessions lengthened, dropping from 17% to 5% detection after 50+ prior prompts, while auto mode's block rate remained constant through the process. Yeah, because I'm tired of hitting enter all the time.

[06:37] Justin: Yeah, exactly.

[06:39] Matt: Production session analysts from May to June showed manual approval sessions contain serious unintended harm at production severity levels more than twice as often as auto mode sessions. Third-party red teaming with Apollo Research reduced the classifier's miss rate on adversarial attacks from 12% to 7% after a hardening cycle. And in prompt injection testing by Directory Labs across 720 attempts, Claude models running auto mode had a 0% attack success rate compared to 5.83% for GPT-5.6-SOL in Codex auto review mode and 19% in full access mode. I am a big fan of auto mode personally, if I know what it's doing, but it is also a way for you to end up having burning a lot of tokens. So if you are, you know, the, the agent goes off on and you've given enough instruction that it can make a bunch of decisions on your behalf. And then all of a sudden it's gone down a path you didn't mean for it to go. And now all of a sudden sudden you burned, you know, 100,000 tokens on something you didn't want it to do because it, it said I was in auto mode. And it, I see sometimes it will like come to a point where it's, it's giving you data, like facts, but then like it's not asking you a question, so it doesn't pause for you.

[07:41] Matt Kohn: Mm-hmm.

[07:41] Matt: But it just, you know, it's like, oh wait, wait, wait, that fact is something I need to know. And now I need to now change what I want you to do next based on the thing you learned. So there are some things about auto mode I don't like. And so it depends on what I'm doing with it or if I'll turn it on or off. But my default going to auto, I don't, I'm not too upset about, to be honest. Yeah.

[07:57] Justin: No, this is, you know, a debate I've lost because I don't think that, because in enterprise controls, you can disable auto mode for your entire enterprise. And I just don't believe in that setting just because I do think that fatigue is worse, right? I think I would rather have people do auto mode and be prompted when it is something risky rather than having to approve everything. 'Cause I, you know, I'm human just like everyone else. And I'll tell you, if I have to sit there and hit yes every 15 seconds, I am not reading that prompt. I just want it to go after a little while. And it's, that's, I really like that they advertise the statistics, which, you know, I totally plan to use and weaponize.

[08:36] Matt: As you should. It's good data. I mean, I think, again, I think it's, you know, the design of it was to, you know, get away from turning auto, you know, allow anything, which was the other mode you had is either prompt me all the time or just let you go do whatever the hell I wanted you to do. Right. You know, without any prompting at all. And that's how people were wiping out laptops and doing all kinds of craziness. So I think this is a good middle ground compromise. I've not tried the Cowork version of it. I, I didn't actually know that existed. It's one of my big complaints with Cowork is that it is super naggy about permissions for everything.

[09:06] Justin: Mm-hmm.

[09:07] Matt: And so I, I'll go check that out with, uh, Cowork. That's one of the re— or sorry, with the, whatever, uh, ChatGPT solution. Codex. Oh, so many tools. Yeah. I will try that out and report back. Yeah.

[09:19] Justin: Between that and then Apple's own application security controls, right? Right. Cowork wants access to your directory. Like the directory I told it to work.

[09:28] Matt: Oh my God.

[09:29] Ryan Lucas: Yeah.

[09:29] Justin: Yeah. Yeah.

[09:30] Matt: It's super annoying. Anthropic has also extended its compliance API to cover Claude Cowork. And Claude Code, currently in beta for Claude Enterprise customers, unifying session visibility with existing Claude Chat coverage. This one needs the applause, Matt. Oh, session records, uh, combines content, prompts, responses, tool calls, skills, artifacts, and metadata into a single consolidated transcript, simplifying audit and e-discovery workflows.

[09:56] Ryan Lucas: Yeah.

[10:02] Justin: This should have been there with launch. I can't believe that they didn't have this and that it was so disparate between their three, their three sort of offerings in a single app. Like it's very confusing and, and I can't believe this wasn't there. So I'm happy to see this.

[10:16] Matt Kohn: You know, I can kind of get Cowork, it wasn't there, but Claude Code, which is like, I feel like the way most, like how these tools got into enterprises in the first place felt crazy that it wasn't there.

[10:28] Matt: So Claude Code had, you know, some abilities, like you could configure sandbox modes and things with Cloud Cowork, and then, you know, inside the sandbox you could— sorry, with inside a code you could configure different things and prevent certain things from happening. But the problem with Cowork was it's basically an agent running on your laptop that can do anything, and it didn't have these compliance things. You couldn't audit what it was doing. Was a major risk for companies who are in SOC compliance or ISO compliance or even HIPAA or PCI. So the fact that this is now available is just a huge benefit because now you can do full audit, full trace ability. And now a lot of the complaints that I was hearing, uh, at other places around cowork concerns, I think this mutes a lot of them, which is great.

[11:08] Justin: Yeah, I think it'll— if nothing else, it'll make people feel more comfortable, and so they won't have just a knee-jerk reaction to no. Yep.

[11:17] Matt Kohn: But no is such an easy answer.

[11:19] Justin: It sure is.

[11:21] Matt: It's— it is not the answer that you need when your CEO is telling you do more with AI. So it's, uh, it's the It's a curse. So Anthropic is doing something I'm not sure about. They're going to be embedding watermarks in AI-generated outputs, citing EU regulatory requirements as the driver of the effort. The stated goal is to help trace the ancestry of AI-generated content, addressing concerns about distinguishing AI output from human-created work. This move reflects a broader trend of AI vendors adjusting product behavior specifically to meet EU AI Act or regulatory, regulatory, uh, reg— related regulatory compliance requirements. Say that 4 times fast. Watermarking approaches vary in robustness, and technical questions remain about whether these methods can be stripped or evaded. This is worth discussion given the compliance framing of it. Development is part of a pattern of AI companies making policy announcements tied directly to regulatory pressure rather than purely technical or user-driven motivations. So the thing I don't know, and why, is how does this show up? Is like every text output now from the model is going to say written by Claude Opus 5 or Claude Sonnet, and that's gonna be in every copy paste and I'm gonna remove it. Or is it getting put into metadata somewhere? Like, how is this being produced in a way for text that makes any sense? Like, and the thing, like, I know Matt was saying earlier, like, well, what about images? I'm like, well, images can't be made by Claude, but you know, at least for most AI images that are being generated by Gemini, they do put in metadata that they were created by Gemini or they were created by ChatGPT. Um, so like for certain things, okay, cool. Like I get it, it makes sense. Like if I'm doing a PowerPoint presentation with Claude and you wanna add to the metadata of the, of the XML for the PowerPoint file that you, you know, you generated some of the text, okay, cool. But then like, how does it know which text was generated by it versus what? Like there's so many questions I have about traceability of this and it feels like, you know, the EU in its infinite wisdom, uh, you know, has asked for something and they're just trying to solve it but not really thinking through the impact people who are trying to actually use this thing day to day. And then also, I mean, it already shows up on all my, on my, uh, pull command, you know, my PR comments that, you know, was authored with Cowork or whatever, which is fine. I don't have a problem with that. But how are you gonna put it into my source code? Are you gonna do that too? Because that's also ugly.

[13:33] Matt Kohn: I assume it's gonna be like the Microsoft— I think it was the Teams or whatever. Every time you would copy, it would say, you know, like have like the name who did it. And every time it's gonna be like that really annoying feature that you have to X out, delete back.

[13:46] Matt: 'Cause I've never seen that feature. Is that must be something you have to turn on.

[13:50] Justin: Yeah, I haven't seen that either.

[13:50] Matt Kohn: It wasn't, it was some tool that every time I would copy from, it would like pull like, like, like from Teams or from whatever the tool it was. And it used to drive me up the wall 'cause every time I would copy, I would have to delete it out. And that's the way I envisioned this where it's just gonna be copy, you're gonna delete out and this is, You know, the way I originally thought about it was Anthropic just checking the box. Like, we are now saying that this is done by it and it's up to the end user not to delete it. But is that really useful or not? Just another story.

[14:21] Matt: Yeah. I mean, I, there was, there was a website, I don't remember which one it was, but it used to be if you copy-pasted text from the website and pasted it, it would say, it would link you back to where it was copy and pasted from, which was cool. Like almost like a, an attribution. And that was a, that was a bit of JavaScript magic they were doing, which was actually a really cool hack. Like to actually, like, I don't remember who that was, but if you actually go look at how they did that in the JavaScript code, it's, it's a pretty clever, uh, use of the DOM. But, uh, yeah, I, again, I don't know how this looks. So like what they've announced, I'm not a big fan of it. How it actually gets presented and how we, what we see out of it, maybe it makes me less concerned, but time will tell.

[14:59] Justin: Yeah, it'll be in the details for sure.

[15:03] Matt: Uh, Meta is releasing MuseGlimmer, a 30 billion parameter open agentic model under Apache 2.0 license designed to run locally on a single consumer GPU rather than requiring cloud infrastructure. Weights are available now on Hugging Face with framework integrations for LLM.cpp, MLX, and ExecuTorch coming soon. The model uses quantization to compress from over 55GB at full precision down to under 20GB, fitting within a 24 to 32GB memory envelope alongside its KV cache and perception encoder. Meta reports minimal degradation on agentic tasks from this compression. MuseGlimmer ships with the speculative decoding Drafter model based on D2F Flash, which proposes multiple tokens at once for the main model to verify in parallel, improving generation speed without changing output qualities. The training combined LoJIT distillation from a larger teacher model, MuseSpark, mid-training on agent-heavy data, and post-training with supervised fine-tuning plus reinforcement learning across coding, reasoning, and agentic domains. Mark Zuckerberg then said, uh, basically that, uh, he wrote an essay basically saying the open weight distribution as a check against AI power concentrating in large institutions, positioning Meta's approach against the closed models used by OpenAI, Anthropic, and Google. Meta defended distillation, the technique where models learn from outputs of other models, calling it a legitimate practice rather than harmful in contrast to accusations OpenAI and Anthropic have made against Chinese AI labs like Deepseek. And Meta announced a $1 billion fund for communities hosting its US data centers, addressing local pushback over power and water resource competition. But in general, uh, the market didn't really like this because they're, they're all over the place. They had Lambda 3 and then they went closed. Now they're back to open. And what's next? Are they going to go back to closed when this model doesn't work? And we talked about it actually a few weeks ago. I don't remember who was here with me, but you know, they're really the only open models that are having any credibility is Deepseek and GLM and KIMI K3. And those are the ones that everyone talks about with open models. Mistral comes up occasionally. Cohere comes up occasionally, but those are, I don't remember if those are open, closed, what exactly, but like they're, you know, they're definitely not the big three. And I, I think I made the comment on the show that Meta kind of missed this because they went closed source, that they kind of were the only US company who was really producing an open model of any size or, or stature other than, you know, Maya models from Microsoft, which they admit are small models. So I think it's a good move. But I mean, man, Zuckerberg, come on. Like no one told you this is a bad idea when you went closed the first time. Like, I don't, it just doesn't make any sense.

[17:30] Matt Kohn: It just feels like it's, you know, multiple failures and it's like, hey guys, you all are doing it wrong because, you know, and it's a way for them, for him to get marketing, I feel like for it more than anything. Like, hey, this is the problem, but you were there last week.

[17:47] Matt: You know, so it just feels like, uh, well, and I think, I think the market basically yawned when he went closed source. They were like, well, it was interesting with Lambda 3. Now if you're closed source, I'm not interested in your new models. And so now, you know, maybe they're reacting to the fact that no one cared about their closed model.

[18:04] Justin: I think the real crime here is that I kind of agree with some of the stuff he said, not all of it. And if it wasn't positioned in a way where they were sort of reacting from what seems like poor market performance, I think it would have a much better ring to it. But it's, you know, there is a consolidation of AI across a very small footprint of companies, and I do like the idea of more open models that you can run on your own hardware. And I was really sad when they went closed source. And so like, this is pretty funny to me that the reverse on this and then how the news is sort of tainted for me.

[18:39] Matt: I mean, I think if I read, you know, if you read the actual essay that he sent, I sort of, sort of felt like he was describing an AI utopia that is going to be terrible for everybody but the rich people. So I was a little, I was a little put off by parts of his essay and like, like this is what everyone's saying they don't want in AI in the future. And yeah, so again, it's one of their tech bro kind of not really understanding what people really want, but. I do agree with his stance on we need more options than just 3 closed source models that are US bound. Because I think there's, I think that's a market that was lacking that they were ceding to the Chinese. Now distillation as a concept, I don't have a problem with either, as long as you're not using, you know, Claude models to build your distillation for yours. But if you create a big foundational model that you own, which they do, and then you distill small models from that, I don't think that's a bad method. So I don't disagree with him on that part either.

[19:35] Justin: Yeah. But I think when he was defending Chinese AI developers, I think it was the former, right? Where it was like, you know, using OpenAI.

[19:45] Matt: Well, that was the claim. Deepseek was that, you know, they used OpenAI to create distillation for Deepseek. That's, that was the original claim. I don't know if that's been proven in court or if that's ever moved anywhere beyond people claiming that was the case. But you know, I definitely remember that was the big fear when they first came out. Yeah. All right. Moving on to AWS, who I guess, what, has a lot of AI stories for us.

[20:07] Justin: No way. Yep.

[20:08] Matt: No way. First up, AWS Bedrock is adding two new gateway-level controls, temporal policies for routing sequences of agent actions and rate limiting to cap token requests and connection consumption per user. Both are available now and require no changes to agent code or existing production deployments. The temporal policies address a gap in traditional authorization models, which check each action in isolation. AgentCore now tracks the state across a session, blocking things like a purchase that would push cumulative spend over budget, even if each individual transaction is under the threshold. These policies are powered by Dogwood, a new open-source policy language built on Cedar, released under Apache 2.0, enforce what happens at the gateway layer outside the agent's own code, meaning the agent cannot reason around or bypass the restrictions regardless of the prompting. RAG limiting lets platform teams set per-second and per-minute ceilings on requests, tokens, and connection duration tied to existing OAuth or IAM identities. And also responds to industry data showing cost and security concerns as a top barrier to Anthropic AI adoption at scale. Yeah, if you wanted to control outputs and token budgets, thanks, you were doing it at the model call itself previously with a number of output tokens required of that. And so, yeah, models would see that, oh, I'm being capped at at 8,000 characters on an output, I can just make multiple requests and then chain them back together. And so, you know, models that are breaking out, uh, of their controls need controls that are outside the model itself. So this is a good move.

[21:29] Matt Kohn: Yeah.

[21:30] Justin: And I do like the, you know, sort of the introduction that they're using here of how they're evaluating the AI responses, which I think is neat, like the new, which we're about to talk about next.

[21:41] Matt: Yeah. So, I mean, Segway right into Dogwood then. So Dogwood, uh, is an open-source governance language that adds temporal sequence-aware policy checks for AI agent tools calls, extending the existing Cedar policy language, which we're a big fan of here on the show. Cedar policies only evaluate a single request in isolation, meaning they can't enforce rules like require approval before selling or limit transfers per hour. Dogwood adds temporal clauses that look back at prior events within a time window, enabling rate limits, order constraints, and running totals. Dogwood is fully backward compatible with Cedar, so existing Cedar policies work unchanged and don't require migration, and teams can incrementally add temporal conditions like formerly, count within, count distinct within, and sum within alongside their current authorization rules. The language is built on metric first order temporal logic, a formal method approach for runtime verification, giving it a mathematical rigor for expressing prerequisites, rate limits, and sequencing constraints on agent behavior. And some re:Invent talk will go deep into this math and I will understand none of it, but I will still watch trying to hope that it gets through my thick skull someday.

[22:43] Justin: Yeah. Yeah. Much like all of their sort of mathematical computation for evaluation. Yeah. I mean, I do, I think this is really neat and addresses a pretty big limitation. And the fact that it's an open source language also means that this can be applied and we'll, you'll see it pop up in other other software components. And so it's like the evaluation is more than just a, you know, a rule-based point-in-time evaluation, which is really neat. You, in order to do this before, which is Cedar, you would have to sort of orchestrate sort of your, your chain of signal in order to write a rule saying, if this doesn't exist as part of this artifact, then don't allow it. But you had to figure out the before part where that would put that in place. Now you can actually just define it as a rule within Docker that this thing have, this action needed to have happened sometime in the future or the past, and then it can evaluate on that, which I think is really neat and a big difference from what we had before.

[23:43] Matt: Well, I mean, and this is an evolution of what we need to do now with identities and agentic identities in particular, 'cause Cedar was written very much in the, I need to verify this user role or user IAM permission has the ability to do these based on a certain set of conditions. and it's a one-time action. But, you know, again, with a GenTech AI and LLMs, it's, it's a sequence of things put together. And so it's not just a single transaction. So you had to evolve, which I, I'm, I'm glad they open sourced it as well. You know, we have seen other team, you know, other companies adopt Cedar. We've seen different implementations of Cedar concepts and different solutions. Like I think StrongDM, for example, uses Cedar pretty heavily in their implementation. And so that's great to see.

[24:21] Justin: Yeah. You know, I, I, I'm pretty happy. To see this, um, just because, uh, you know, having used Cedar, like, it can be really complicated. And yeah, you know, before we were trying to embed, you know, in any kind of agentic workflow on behalf of, for triggering something that you needed to use basically the user permissions, but evaluating that, like, and if it wasn't properly metadata tagged, then you just had to deny it. This way you can look at a group of actions and make that determination. Which is awesome. Yep.

[24:52] Matt: AWS added runtime instances to Bedrock Agent Core runtime, giving developers dedicated EC2-backed infrastructure for AI agents that need to run for multiple days, access GPUs, or coordinate with other agents on the same host, complementing the existing microVM option capped at the 8-hour invocation. Sessions persist for up to 14 days with stop restart support, so long-running workflows can hibernate over a weekend and resume without losing state, which addresses a common pain point for multi-day agent tasks like code generation and review pipelines. Multiple agents can be deployed on a single runtime and collaborate through a shared file system within a session, as demonstrated with a code writer and code reviewer agent exchanging work without any direct API calls between them. Setup requires creating a capacity provider that defines the EC2 instance type, OS, VPC, and storage, then deploying agents via S3 upload or container image with support for any framework and any model. It's 14 days because on the 15th day, it comes to the singularity and destroys the world. So 14 days is the arbitrary date where that does not happen. Uh, is the only reason why I can think it's 14 days. Yeah.

[25:52] Justin: It is kind of a weird number. So wait, not everyone's just doing Markdown files to the end of time to maintain their state across sessions? Like, oh, I never really thought about persisting them. Like, how crazy.

[26:05] Matt: Right? Uh, someone, there's some tool, like, it's not Notion, it's an open source tool, and people have raved about Obsidian. Obsidian. Thank you.

[26:12] Justin: Yeah. Yeah.

[26:13] Matt: You know, they rave about like, oh, use Obsidian as your like stored memory place where the agent can go look up things. And I have not tried that. So if either of you have done it, I would love to hear your feedback on that because it's, it's on my list of like, I should try that someday, but then I'm like, I don't need another text editor.

[26:27] Justin: Yeah, the, the promise of it is really, uh, really cool, but I've also seen stuff that I think is a little, a little bit more advanced where it's actually using vector search and data databases to maintain that instead of having, yeah, another text editor.

[26:44] Matt Kohn: I just, do the markdown files. Maybe I'm overly simplistic, but it works right now for me. Yeah.

[26:51] Justin: I mean, I bet it works great. I've read it and the promise of it looks amazing. Trying to get it into my workflow has been the challenge point for me, which is like, I'm already using too many apps for AI stuff. Like, you know, cause I'm using multiple different developer patterns. And then there's also like everyone has their own chat interface now. Like even Atlassian's AI bot has their own app. And so I just run that locally now cuz it's, you know, so much better than trying to figure out Jira query language.

[27:18] Matt Kohn: I didn't realize they released a full app for it. Yeah, it's like a whole Rovo, or is it Rovo app?

[27:23] Justin: Whole Rovo app, man. Yeah, you can just chat back, chat, add it up.

[27:27] Matt Kohn: Yeah, I missed that one.

[27:29] Matt: I've seen Rovo in the, in the SaaS console.

[27:32] Matt Kohn: Yeah.

[27:32] Matt: But I didn't know that you could get a client for it.

[27:34] Justin: That's nice. You can get a full client. And so now I just ask a question.

[27:36] Matt Kohn: I just use Claude with the MCP and let it deal with that in the background.

[27:41] Matt: I've done that. I've done that too. Yeah.

[27:42] Justin: Yeah.

[27:42] Matt Kohn: Yeah.

[27:43] Justin: That's what I was doing before the robo thing. I did, I had found the robo thing because I have to evaluate all this stuff for the day job. And I was like, oh, and it works really well for finding stuff. Like, I suspect that it's because it has access to things that, and is providing data that I don't have access to on the backend. But, you know, if I, so far I haven't been able to prove that. It just feels That's true.

[28:05] Matt Kohn: You're like, it works too well. Yeah.

[28:10] Matt: DynamoDB is now apparently supporting native vector search, letting customers store embeddings alongside operational data and run similarity searches without maintaining a separate vector database or sync pipeline. So I mean that whole S3 vector thing, I guess we didn't need.

[28:23] Matt Kohn: So there's S3 vectors, there's DynamoDB vectors, there's Redis vectors. Redis. Yep. Doesn't OpenSearch have vectors in it now?

[28:32] Matt: Yes, it does.

[28:34] Justin: I think you can do vectors in Aurora now too, for Postgres.

[28:37] Matt Kohn: Postgres has pgVectors. Yeah.

[28:40] Matt: And SQL Server has them.

[28:41] Matt Kohn: So I'm just saying all databases count as 6, number 6.

[28:44] Ryan Lucas: Yeah.

[28:44] Matt: I mean, basically just vectors are a thing a database is gonna support.

[28:47] Matt Kohn: Yes. There was one other one I thought they announced a few weeks ago that I can't remember, but yeah. Anyway, sorry.

[28:53] Justin: Is that gonna be our rule now? You know, a data project doesn't become real unless it inputs SQL. Now it's gonna be a database doesn't become real until it as a vector, sure.

[29:03] Matt: Yeah, I mean, NoSQL wasn't a database, still had an index, so yeah, there we go. Yeah, maybe that's the new one. We'll take you behind that. But hey, it's good to know. I'm still using S3 vectors for Bolt, which is good enough. Most of our problem with Bolt right now is just embeddings are too big. I need to redo the embeddings, make them smaller chunks. But yeah, it's just been not something I've had time to do in my busy time.

[29:27] Justin: DynamoDB would force that issue, right, with the, the index type limits.

[29:31] Matt: It would, yes, yes it would. Amazon Cognito is now available as the AWS Auth skill in the agent toolkit for AWS, letting AI coding agents set up, configure, secure, and troubleshoot Cognito using pre-built best practice workflows instead of manual configurations. The skill covers a broad scope of Cognito functionality, including user pools, app clients, OAuth 2.0 flows, JWT authorizers, passkey, WebAuth, enrollment, threat protection, Lambda triggers, and identity pools. When paired with the AWS GCP server, agents execute AWS CLI commands with IAM-based guardrails and CloudTrail audit logs. Now the true question is, will Cognito force the AI to go full Terminator on us, Skynet? Or will this prevent them from actually being successful and this is the best thing Amazon could ever release? It can only go one of two ways.

[30:16] Justin: Oh, I hadn't thought of the, the alternative, like, you know, 'Cause I was too busy like, well, maybe AI will have a chance of figuring out Cognito and then realizing that no way is that gonna work where you can actually guarantee that you can get into your application.

[30:30] Matt Kohn: But that's why it can't prevent it. Yeah, no, no, no. AWS uses it for so many things. Terminator's not gonna be able to get to anything, so we'll be good.

[30:41] Justin: Yeah, access denied for the apocalypse. I like it.

[30:45] Matt: Yep.

[30:46] Justin: That's what we should have used for the show title.

[30:48] Matt: There you go. That's what we should have done.

[30:51] Matt Kohn: After like 20 minutes of trying to figure that out, that's what we should have done. We should really do show titles at the end of the show, I feel like.

[30:57] Justin: Yeah. Yeah.

[30:59] Matt: I mean, other shows do it that way. We, we can definitely start doing that. It'd just be more work for editing.

[31:04] Justin: Yeah.

[31:05] Matt: In the category I like to call, thanks Nova, Amazon has released application status checks as a, for EC2 to close a longstanding gap by monitoring actual application health not just instance and system reachability, catching issues like a stopped web server, a crashed Docker daemon, or misconfigured networking that previous status checks would have missed. Setup is straightforward, with customers defining a protocol, port, and path, along with expected healthy response codes, then associate the check with instance by ID or tag. ECS will pull this every 60 seconds and report the status. Integration with Auto Scaling groups means unhealthy applications can trigger automatic instance replacement, reducing the need for custom health check tooling that many teams previously built and maintained themselves. Yes, please raise your hand if you've done it. All of us on the podcast.

[31:47] Justin: Yep.

[31:48] Matt: Availability spans all commercial AWS regions plus the GovCloud. So this isn't a limited preview. Thank God. And thank you, Nova. I really appreciate you writing that out for me.

[31:58] Matt Kohn: So I both like and dislike this.

[32:01] Justin: Oh, dislike. Okay, go on.

[32:03] Matt Kohn: I dislike that it's associated directly with EC2 instance. Like, I don't mind the Auto Scaling Group and I like that it's associated with it, And I get why it's associated with the EC2 instance, a one-off, like OpenVPN or something like that. Like, but I just feel like this is going to be a really— I say this knowing that I'm going to be the cause of this. Like, I'm going to use this to hack around something that shouldn't be done. Like, it just feels like this is going to open up like bad Pandora's boxes.

[32:34] Justin: That's funny because you and I have absolutely in container orchestration created this tool in a very bad way using Application Load Balancer and dummy, basically health check. Yeah. And so like, and what we needed was this, and it was specifically for stuff like a Docker daemon that stopped responding or, you know, or ECS agent or, you know, different things like that.

[32:56] Matt: Right.

[32:56] Justin: And so it's like, if you had this natively, we could have just used, it would have been a much simpler rollout than, than what we had.

[33:03] Matt: I mean, are people gonna abuse it? Yeah, so they abuse health checks today too.

[33:06] Justin: Yeah.

[33:07] Matt Kohn: Yeah. But I just, I kind of wish it was just with the auto scaling group just because then like, someone's gonna set it up with the EC2 instance and like have this trigger like a Lambda that just terminates their EC2 instance, which I guess if you do something stupid like that, 'cause I'm trying to think of something on the fly, you deserve to lose your data. Yeah.

[33:25] Matt: But I say like, this sounds like you know, you were properly scolded by infrastructure.

[33:32] Matt Kohn: But like, it just feels like someone's gonna do something really stupid because this is here, or they're gonna configure it, it's gonna auto-stop, and you're gonna end up with like the boot loop cycle where like it boots up and shuts down. Like, I'm sure you guys have come across this on EC2 in auto-scaling group before too, where like your health check isn't set right and it goes in a boot loop. Like I just, I both like and hate it. Like I get why it's here. I fully appreciate it's here. It will make my life easier. I just am not in love with it yet.

[34:03] Justin: I just don't know. So it's attached to the EC2 host, but it's, unless you attach it to like a Lambda that's looking for an event on EventBridge or it's just gonna error out, I guess, or you're an auto scaling group, nothing's gonna happen. So I don't, I don't understand the harm in it.

[34:18] Matt: I don't know. He's worried about future Matt who's gonna bastardize it for some terrible purpose.

[34:23] Justin: Future Matt would've done something else terrible.

[34:26] Matt: Yeah. You would come up with a worse solution. So now at least, at least it'll at least be somewhat based on a cloud native idea versus whatever Rube Goldberg machine you came up with.

[34:35] Matt Kohn: Yeah.

[34:35] Justin: I'll have half a chance of being like, oh, I think he just did this and then being able to undo it. Yeah.

[34:40] Matt Kohn: So is this the fourth now also alert that's on the EC2? Cause there was the original two, they added a third. And then there's the fourth.

[34:48] Matt: This is now the fourth one. Yep.

[34:49] Ryan Lucas: Oh, really?

[34:51] Matt: Yeah. But the third one was the, was basically making sure that the cloud init basically launched properly and completed.

[34:59] Justin: Yeah.

[34:59] Matt: Secrets Manager now handles automatic rotation for Jenkins API tokens and SonarQube tokens without custom rotation code, extending its managed external secrets list to 11 supported third-party services, including GitLab, Okta, and Snowflake. Jenkins rotation uses a verify before revoke approach, minting a new token and confirming it works before killing the old one, which avoids CI/CD pipeline interruptions during credential swaps. I mean, I prefer breaking CI/CD. Both self-rotation and admin-assisted rotation are supported depending on how teams manage token permissions. SonarQube integrates, uh, covers 3 token types, user, global analysis, and project analysis, with user tokens supporting self-rotation and analysis tokens requiring an admin token for rotation. This targets a common pain point for DevOps teams. Manually rotating CI/CD and code quality tool credentials is tedious and often skipped. Leaving long-lived tokens as a security risk.

[35:47] Justin: Well, the fact that this supports Salesforce's external client secret, I think is a, given, you know, some of the very public breaches lately, I'm like, yeah, this is a good idea. We should do that. Mm-hmm. Yeah. So external secrets and auto rotation is awesome.

[36:00] Matt: You know, when they first rolled this out, I was, I was excited about it and I was hoping more companies would adopt it faster. And my only complaint is that more companies aren't rolling this capability out.

[36:10] Justin: And of those 11, 3 of them are Datadog.

[36:13] Matt: Great.

[36:14] Matt Kohn: Yeah.

[36:15] Matt: Which is a commentary on Datadog's integration in general. But yeah.

[36:20] Justin: 3 different credential management platform. That sounds good.

[36:23] Matt: Yeah. All right. GCP is introducing America's Connect, which is, you guessed it, subsea cables. The best. These are 3 new cable systems, the Alicios, which will connect to Dominican Republic and Panama and Chile. The Kanoa, which will be Dom— uh, the Dominican Republic to Bermuda. And the Ololas, I guess that's— if I butchered that, I'm sorry— Dominican Republic to Florida. Plus a new Fermina branch landing in the Dominican Republic as well. Fermina is one of the ones we talked about here previously. The combined network creates redundant ring topology routes across the Pacific Coast, Caribbean Sea, and Atlantic Ocean, directly connecting Google Cloud regions in Chile, Los Angeles, Las Vegas, South Carolina, Virginia, and Madrid. The expansion builds on prior investments in the Curie, Nuvem, and Sol cables, positioning the Dominican Republic as a central hub linking Latin America, the Caribbean, North America, and Europe. The approach mirrors Google's specific Connect initiative, using strategically placed branching units to allow future expansion as regional connectivity needs are growing. Uh, super cool.

[37:23] Justin: Yeah, and illustrates my lack of geography when I always forget that South America and Central America so much further east than I think they are.

[37:33] Matt Kohn: They're not directly below, they're off.

[37:35] Justin: Yeah, no, yeah, it's a globe, it's round, it's, it's round and it, it, it, it curves and the map is a lie.

[37:44] Matt: Yeah.

[37:44] Matt Kohn: Did you guys look at the, the image they have there? Hurts my brain. I can't figure out like the lines, the color, like they chose very weird ways to denote these. They're like—

[37:55] Matt: I mean, this was, this was clearly designed by a colorblind person.

[37:58] Matt Kohn: They've always, I'm colorblind and all of their announcements have been this same format.

[38:02] Matt: So it's pretty funny. They like this format and it's always hard to read.

[38:05] Matt Kohn: Yeah. I don't understand it, but it's also interesting. Dominican Republic is gonna be like, kind of like a centralized hub for all their stuff, which is, you know, not where I would've expected.

[38:15] Justin: No, exactly. You know, it's, it, but you know, cool. Good for them. I mean, it's good for the industry. I think that region could use it.

[38:22] Matt: They just tells the Russians where to attack if they wanna take down the Caribbean, I guess.

[38:26] Matt Kohn: In some ways you just take out Bermuda.

[38:29] Justin: Yeah.

[38:30] Matt: One of those will work. So, yeah, but I'm, I'm glad to see it though. I think it's super cool. And, uh, we, I still need to reach out to some cable companies cuz it's, it's a bit of a tough industry to crack into. I tried to go through Google and I couldn't get anybody at Google to come talk to us, but I, I am, I do, there's a couple other vendors you need to reach out to. I started working on it and then I got busy. Uh, I think it's one of those like secret things, like, you know, you see there's, I think there's, I think even if we do an interview, it's gonna be highly redacted what they can talk about and they can't talk about. And so we'll have probably pretty clear guidelines because again, it is a national security concern in some ways as well. And so it'll be, it'll be a mix. And then on to Azure. Microsoft is finally deprecating SMS as authentication method in Entra ID, pushing passkeys as the preferred default with a key rollout deadline of September 1st for organizations to prepare. You know, in general, uh, this is probably not the best week to announce a lot more PASKY stuff because there was like a PASKY exploit that happened in this last week that I saw. But, uh, you know, getting rid of SMS is definitely a good idea. If you've ever been to a third world country where they have smishing and all kinds of other cell phone, uh, you know, attack vectors, they're, they're like blasting different radio signals to try to get your cell phone to connect and steal your data. Uh, yeah, SMS is a bad choice. You should not use it.

[39:48] Matt Kohn: Well, GitHub wouldn't even let you, like, if you had like MFA, like, like secure MFA on it, I don't remember what they called it, it would actually not let your account be tied in to your org if you still had, um, SMS turned on.

[40:02] Justin: Yeah.

[40:02] Matt Kohn: Because I had a whole conversation with somebody, they're like, I have two-factor, it's text. And I was like, text isn't secure anymore. So no, you gotta adjust that.

[40:12] Matt: Can we talk about though, the fact that if you want to use passkeys with Microsoft, you have to use Microsoft Authenticator and how much I hate Microsoft Authenticator.

[40:21] Matt Kohn: You have to unlock your phone and then unlock your app. And then if you have to go do something else, re-unlock the app again a second time.

[40:28] Matt: Yeah. Well, but you unlock the app and then it sends the push notification to the app, which is then you have to Face ID again.

[40:34] Justin: Right. Yeah. Right.

[40:35] Matt: So it's like, it's like 3 Face IDs by the time you get to an actual point where a thing works.

[40:40] Justin: Which is a long time for me to stare at my phone. 'Cause if I look away for a second, it breaks it, right? Like, yeah.

[40:46] Matt: And then I was like, why didn't this log in? Oh, I forgot the third, the third time I had to stare at my phone for a second.

[40:51] Matt Kohn: Mm-hmm. My problem is I get distracted by some popup or something else on my phone and my ADD kicks in and I'm like, ooh, shiny objects over here.

[40:59] Matt: You know, there's also the fact that, you know, every time you have Authenticator and you have a, you know, a corporate management profile, they take over Authentica— they take over Authenticator. As one of the apps they want to control, which then causes you all kinds of havoc when you leave a job and they revoke your MDM and then your personal stuff that's in Authenticator is now also gone.

[41:18] Justin: Yeah.

[41:18] Matt: So there's lots of fun ways that Authenticator burns you that just really annoy me. So I just decided I will never use Authenticator for anything but Microsoft products, uh, because you have to, unfortunately, because they don't let you use third parties, which is also annoying.

[41:31] Justin: Yeah.

[41:32] Matt Kohn: Yeah, I only use it for Microsoft at this point. Like it's just easier.

[41:37] Matt: Yeah. So, uh, I, I just wish they would let me use whatever I want to use for passkeys. It'd be nice versus their one true way, apparently.

[41:46] Justin: Yeah.

[41:47] Matt Kohn: They know better than you, Justin. They know security. You don't.

[41:51] Matt: I don't know that I believe you on that.

[41:54] Justin: Yeah.

[41:55] Matt Kohn: I tried with such a straight face.

[41:57] Justin: I couldn't keep it up.

[41:57] Matt: You tried. You couldn't keep it up.

[41:59] Matt Kohn: I couldn't.

[41:59] Ryan Lucas: Yeah.

[42:01] Matt: And then, uh, our last story for the week, uh, in a cloud story that's not very cloudy, Azure Virtual Network Routing Appliance is now generally available, offering dedicated hardware for east-west traffic routing between virtual networks instead of relying on VM-based solutions with bandwidth tiers up to 200 gigabits per second per instance. The appliance supports IPv4, IPv6, and dual-stack configurations at scale, including IPv6 access control list enforcement, which addresses a gap for organizations managing complex multi-region network topologies. Key operational benefits is the fully managed nature of the service, including built-in high availability and availability zone resiliency, removing the maintenance burden compared to self-managed VM-based routing appliances. So I have a question, Matt. Where would I install this appliance in my Azure clouds?

[42:48] Matt Kohn: I plead the fifth.

[42:52] Matt: Because that's where I'm like, I'm just perplexed. I'm like, okay, A, why wouldn't the cloud just provide this to me as a virtual appliance or something I just configure as part of— or just a managed service? Or a managed service.

[43:02] Justin: Yeah.

[43:03] Matt: But like the way they, the way they talk about this is it's hardware device that I have to install. Like, so am I, am I getting an Equinix contract to go put this hardware appliance in to route between my VNets and Azure? Like, I don't understand what they're actually, what this actually is, which is— Yeah.

[43:16] Justin: Do you have to have a lot of cross-VNet traffic?

[43:18] Matt Kohn: Yeah.

[43:18] Justin: Do you, I mean, is it like, or can you just configure like these two VPCs can talk to each other and that's handled in the backend?

[43:24] Matt Kohn: Well, that's the point. This is like a bad version of Transit Gateway.

[43:28] Matt: A really bad version of it.

[43:30] Matt Kohn: Like, and then you have to predefine, and I haven't looked too much into it, but my understanding is you select your bandwidth and like the pricing, like you want 10 gigabits, $0.70 per hour out to start off. You want 200, $13.05 per hour. And I don't even know if that includes redundancy or what it is. And I don't know if you can change. So like, I just feel like all this is bad news bearers across the board. It's somebody like some, some person's like, oh, we need to do this, but we don't want to develop this to be cloud native. Great, we'll just do this. And you know, I get it's designed for hub and spokes and not to egress and to for east-west traffic, but there has to be, you know, they already had some of these things in there. So like this just has to be for massive high bandwidth.

[44:18] Matt: I mean, it does say 200 gigabits per second, which is a pretty big pipe for sure.

[44:23] Justin: Oh yeah.

[44:24] Matt: Yeah.

[44:25] Justin: I mean, I just feel sorry for whoever has the workload who has to use this plus the, you know, the BGP routing policy that they announced last week. Like, oh man, you have the BGP route map, whatever stack needs this.

[44:35] Matt: Oh yeah. There's a lot of companies who have built data centers in the sky. And so that is, uh, how you solve this problem in those worlds. So it's an unfortunate reality, but that's a lot of companies are They're in it.

[44:48] Justin: So, mm-hmm.

[44:49] Matt: Well, gentlemen, we have made it to another end of the podcast.

[44:53] Matt Kohn: Nice and quick today. Good job, guys.

[44:55] Justin: Yeah.

[44:56] Matt: 45 minutes. I mean, it wasn't a, it wasn't a huge news week. Uh, at least that we felt worthwhile talking about. There was some other stuff that we cut, but yeah, it was worth cutting. Worth cutting. You missed nothing.

[45:06] Matt Kohn: Yeah.

[45:06] Justin: And we were just gonna say dumb things about it, you know?

[45:09] Matt: Yeah, probably.

[45:10] Matt Kohn: Yeah.

[45:10] Matt: We always say dumb things about it.

[45:11] Matt Kohn: To be fair, that's why people listen to us.

[45:13] Matt: Yeah. I mean, I love when they come and correct me. That's always my favorite. Like, oh, thanks. Yeah, you're right. I did get that wrong. I appreciate it. So, all right. Well, uh, we will see you next week here in the podcast world.

[45:25] Justin: Bye everybody.

[45:26] Matt Kohn: Bye everyone.

[45:28] Ryan Lucas: Another week of cloud news wrapped up. Boat will collect the news. Justin will get the notes. Jonathan will write some code. Ryan will watch the perimeter. Jonathan and Matt will reluctantly watch Azure till next week for AI, Amazon, Google Cloud, and Azure. And hey, maybe even Oracle, who knows? Check out thecloudpod.net for our newsletter. Join our Slack, message us on socials, or leave a review.