Welcome to episode 367 of The Cloud Pod, where the forecast is always cloudy! Justin, Ryan, and Matthew are in the studio this week and ready with a lot of news, including passkeys (we know, they’ve had a rough week), Secrets Manager, Vector Search, and Glimmer (no, not my second favorite character from She-Ra), and even…wait for it…undersea cable news!
We’ve got a lot to cover, so let’s get started!
Titles we almost went with this week
- AWS Secrets Manager Jenkins Rotation Finally
- Claude Enterprise Hooks a Ride on Data Loss Prevention
- Passkeys Take the Wheel, SMS Rides Off Into the Sunset
- Claude Code Says Trust Falls Are Over
- Muse Glimmer Shines While Meta’s Wallet Dims
- Zuckerberg Bets Big on Open Weights, Loses on Free Cash Flow
- AI is persistently in the news
- How many ways are there to run vector search in AWS, now 1 more
- Vector Search is the new Docker on AWS… how many ways are there to run it
- AWS Says “You get a Vector Search, and you get a Vector Search”
- You say you’re a Cloud Azure, but “Azure Network Router Appliance” says otherwise
- Claude now tells the world, I did the AI Slop
- Open, Closed, Open; Zuckerberg is on the AI Revolving Door
- Anthropic triples everyone’s productivity with Automode
A big thanks to this week’s sponsors:
We’re sponsorless! Want to get your brand, company, or service in front of a very enthusiastic group of cloud news seekers? You’ve come to the right place! Send us an email or hit us up on our Slack channel for more info.
AI Is Going Great – or How ML Makes Money
01:40 Inference hooks: inline data loss prevention for Claude Enterprise
- Anthropic launched inference hooks in beta for Claude Enterprise, providing inline data loss prevention across chat, Claude Code, Claude Cowork, and other Enterprise surfaces through a single configuration point.
- Technical approach: every inference request routes through a signed WebSocket connection to a customer-controlled security server; Claude sends the prompt and context before generation begins and waits for an allow/deny verdict before proceeding. The same inspection applies to tool call responses, including those from MCP connectors, skills, and plugins.
- The feature uses an open, webhook-based protocol with a published schema, allowing integration with existing DLP vendors such as Netskope, Palo Alto Networks, Proofpoint, and Zscaler, or custom in-house security servers, without requiring separate per-product integration work.
- 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.
- This addresses a gap where inline enforcement was previously limited to Claude Code’s client-side hooks, giving compliance teams a unified enforcement layer for sensitive data across all Claude Enterprise channels.
- Documentation is available here.
05:10 Ryan – “Anthropic has their own issues, so you can just blame them every time.”
05:59 Auto mode is now the default in Claude Code for Pro, Max, and Team plans
- Starting August 14, 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 partner 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: 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.
- Production session analysis (May-June 2026) showed manually approved sessions contained serious unintended harm at production-severity levels more than twice as often as auto mode sessions (6.3% versus 2.4%). Third-party red-teaming with Apollo Research reduced the classifier’s miss rate on adversarial attacks from 12% to 7% after a hardening cycle.
- In prompt injection testing by Trajectory 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’s Auto-review mode and 19.03% in Full Access mode, highlighting differences in built-in safeguards between agentic coding tools.
- Adoption data shows practical impact: teams using auto mode ship about 25% more PRs, and case studies from Adobe, Nuro, Gusto, and Garner Health illustrate use cases like overnight autonomous agents, standardized SDLCs without command allowlists, and reduced permission fatigue.
- Anthropic also shared three internal incidents where the classifier prevented data leaks, mass destructive operations, and privilege escalation mismatches.
07:10 Justin – “I am a big fan of automode, personally, if I know what it’s doing; but it is also a way for you to burn a lot of tokens. The agent goes off, and you give it enough instruction that it makes a decision on your behalf, and all of a sudden it’s going down a path you didn’t mean for it to go, and you burned 1000,000 tokens on something you didn’t want it to do.”
09:37 Compliance API coverage extends to Claude Cowork and Claude Code
- Anthropic has extended its Compliance API to cover Claude Cowork (desktop, web, mobile) and Claude Code (CLI and desktop app), currently in beta for Claude Enterprise customers, unifying session visibility with existing Claude chat coverage.
- Each session record combines content (prompts, responses, tool calls, skills, artifacts) and metadata (verified user ID, email, org ID, session/message IDs, timestamps) into a single consolidated transcript, simplifying audit and eDiscovery workflows.
- The update is additive with no breaking changes to existing Compliance API integrations, and organizations already using OpenTelemetry exports can continue doing so alongside the new endpoints without added infrastructure.
- Coverage gaps remain: this beta excludes Claude Code on the web, Claude Code via the Claude Platform, and sessions run on Amazon Bedrock, Google Cloud Vertex AI, or Microsoft Foundry, so multi-cloud deployments won’t get full visibility yet.
- No new integration work is required for enrolled organizations; they can query the new endpoints directly using their existing Compliance Access Key, lowering the operational overhead for security teams monitoring AI usage across surfaces.
- Documentation is available here.
09:48 Justin – “This one needs the applause.”
11:37 Anthropic pledges to embed watermarks to help discern AI slop in sop to EU
- Anthropic will embed watermarks in AI-generated output, citing EU regulatory requirements as the driver for 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 related regulatory compliance requirements.
- Watermarking approaches vary in robustness, and technical questions remain about whether these methods can be stripped or evaded, which is worth discussing given the compliance framing.
- The 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.
15:19 Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device
- Meta released Muse Glimmer, a 30-billion-parameter open agentic model under an 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 llama.cpp, MLX, and ExecuTorch coming soon.
- The model uses quantization to compress from over 55 GB at full precision down to under 20 GB, fitting within a 24-32 GB memory envelope alongside its KV cache and perception encoder. Meta reports minimal degradation on agentic tasks from this compression.
- Muse Glimmer ships with a speculative decoding drafter model based on DFlash, which proposes multiple tokens at once for the main model to verify in parallel, improving generation speed without changing output quality.
- Training combines logit distillation from a larger teacher model (Muse Spark), mid-training on agent-heavy data, and post-training with supervised fine-tuning plus reinforcement learning across coding, reasoning, and agentic domains.
- The model targets local agent use cases, including tool calling, coding, multimodal inputs via screenshots and documents, and multi-step task completion, and is benchmarked against Gemma4-31B and Qwen3.6-27B on tasks such as SWE-Bench and tau-Bench.
- Distribution partners include Ollama, LM Studio, vLLM, Together AI, and OpenRouter, with hardware optimization support from AMD, Arm, Dell, Intel, and NVIDIA.
- Documentation is available here.
16:45 Justin – “In general, the market didn’t really like this, because they’re all over the place. They have Llama 3, then closed, now back to open, and what’s next?”
con’t. Mark Zuckerberg attacks ‘closed’ AI rivals as Meta returns to open models
- Meta released open weights for its new Muse Glimmer AI model, with weights for the more powerful Muse Spark model expected in the coming weeks; this follows Meta’s earlier decision not to release Muse Spark weights due to safety concerns.
- Zuckerberg’s accompanying essay frames 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.
- Meta announced a 1 billion dollar fund for communities hosting its US data centers, addressing local pushback over power and water resource competition as AI infrastructure build-out continues.
- The announcement follows an 8 percent share price drop after Meta disclosed a 91 percent decline in free cash flow due to AI infrastructure spending, highlighting the financial tension between open model strategy and infrastructure costs.
AWS
20:34 Control agent behaviors and cost beyond a single action: new capabilities in Amazon Bedrock AgentCore
- AWS Bedrock AgentCore adds two new gateway-level controls: temporal policies for evaluating sequences of agent actions, and rate limiting to cap token, request, and connection consumption per user.
- Both are available now and require no changes to agent code or existing production deployments.
- Temporal policies address a gap in traditional authorization models, which check each action in isolation.
- AgentCore can now track state across a session, blocking things like a purchase that would push cumulative spend over budget even if each 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.
- Enforcement happens at the gateway layer outside the agent’s own code, meaning the agent cannot reason around or bypass the restrictions regardless of prompting.
- Rate limiting lets platform teams set per-second and per-minute ceilings on requests, tokens, and connection duration, tied to existing OAuth or IAM identities.
- This targets three distinct failure modes: retry loops (request volume), reasoning-heavy tasks (token consumption), and long idle sessions (open connections).
- The announcement responds to industry data showing cost and security concerns as top barriers to agentic AI adoption at scale, per McKinsey and Forrester research cited in the post.
- Positioning these controls in AWS’s managed infrastructure layer rather than application code is meant to reduce the burden on teams building and approving individual agents.
- Pricing information is available here.
22:20 Introducing Dogwood: runtime verification for AI agents
- AWS released Dogwood, an open-source governance language (Apache 2.0) that adds temporal, sequence-aware policy checks for AI agent tool calls, extending the existing Cedar policy language used in Amazon Bedrock AgentCore Policy. Available now on GitHub at github.com/dogwood-policy/dogwood.
- 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, ordering constraints, and running totals.
- Dogwood is fully backward compatible with Cedar, so existing Cedar policies work unchanged and don’t require migration. 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 (MFOTL), a formal methods approach from runtime verification, giving it mathematical rigor for expressing prerequisites, rate limits, and sequencing constraints on agent behavior.
- This matters for use cases like financial transfer limits, requiring approval workflows, and preventing data exfiltration after accessing confidential information.
- Current limitations include no support yet for absolute time windows (e.g., daily quotas resetting at midnight), no liveness checking (verifying required actions eventually happen), and no multi-agent orchestration policies; all listed as planned future work.
- Temporal policies also lack the automated reasoning and analysis tools that Cedar provides, and evaluation cost scales with event log length.
23:26 Ryan- “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… you’ll see it pop up in other software components.”
25:33 Runtime instances: persistent compute for production AI agents on Amazon Bedrock AgentCore
- AWS added runtime instances to Bedrock AgentCore 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 8-hour invocations.
- 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 (CrewAI, LangGraph, LlamaIndex, Strands) and any model.
- Pricing is standard EC2 rates plus an AgentCore orchestration management fee, and the feature is available at launch in nine regions including US East, US West, select Asia Pacific regions, and Europe (Frankfurt, Ireland), with Linux ARM64 and x86_64 support and Python 3.11-14 plus container images.
26:22 Justin – “It’s 14 days because if on the 15th day it becomes the sim singularity and destroys the world.”
28:57 Amazon DynamoDB now supports real-time vector search at any scale
- DynamoDB now supports native vector search, letting customers store embeddings alongside operational data and run similarity searches without maintaining a separate vector database or sync pipeline.
- This eliminates the data movement costs and licensing overhead of dedicated vector stores.
- Performance specs include single-digit millisecond latency at over 99% recall, scaling to trillions of vectors with no storage limits, using the same serverless, pay-per-request pricing model as standard DynamoDB tables.
- Implementation is straightforward: vectors are stored as DynamoDB’s existing List data type (no new data type needed), and a new vector index type supports up to 4096 dimensions with Euclidean, Cosine, or Dot product distance functions plus inline filtering on non-vector attributes.
- Use cases include retrieval-augmented generation, agentic memory, recommendation engines, and anomaly detection, positioning DynamoDB as a direct option for teams building AI applications who want to avoid running a separate vector database like OpenSearch or Pinecone alongside their operational store.
- Generally available now across all commercial AWS regions and GovCloud (US), with embeddings generated via Bedrock Titan Text Embeddings, Cohere Embed, or OpenAI models.
- Pricing follows standard DynamoDB pay-per-request rates.
30:23 Amazon Cognito now available as a skill in the Agent Toolkit for AWS
- 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 configuration.
- The skill covers a broad scope of Cognito functionality: user pools, app clients, OAuth 2.0 flows, JWT authorizers, passkey/WebAuthn enrollment, threat protection, Lambda triggers, and identity pools, addressing both human user and machine-to-machine authentication scenarios.
- When paired with the AWS MCP Server, agents execute AWS CLI commands with IAM-based guardrails and CloudTrail audit logging, adding a governance layer to agent-driven infrastructure changes; the skill also works standalone via the CLI without requiring the MCP Server.
- This reflects AWS’s continued build-out of the Agent Toolkit with specialized service skills, aiming to reduce the time developers spend translating security best practices into working authentication configurations.
- Available now via GitHub (aws/agent-toolkit-for-aws) and the Agent Toolkit Quick Start guide; no additional cost for the skill itself, though standard Cognito pricing (based on monthly active users) still applies.
26:22 Justin – “The true question is, will Cognito force the AI to go full Terminator on 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.”
32:00 Amazon EC2 introduces application status checks
- EC2 application status checks close a long-standing 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 miss.
- Setup is straightforward: customers define a protocol, port, and path along with expected healthy response codes, then associate the check with instances by ID or tag; EC2 polls every 60 seconds and reports 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.
- Availability spans all commercial AWS Regions plus AWS GovCloud (US), so this isn’t a limited preview; it’s broadly accessible from day one.
- Worth discussing on the show: how this compares to existing solutions like ALB health checks or third-party monitoring tools, and whether this reduces the need for services like Route 53 health checks or CloudWatch synthetic canaries in certain scenarios.
- Pricing details are in the EC2 User Guide, worth checking before recommending broad adoption.
- File this under “Thanks, Nova.”
35:57 AWS Secrets Manager adds managed external secrets support for Jenkins and SonarQube
- 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.
- Both self-rotation and admin-assisted rotation are supported depending on how teams manage token permissions.
- SonarQube integration covers three token types (User, Global Analysis, 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.
- Available in all regions where Secrets Manager managed external secrets is supported; pricing follows standard Secrets Manager rates (per secret per month plus API call charges), with no additional cost called out for these new integrations.
- Check out the documentation here.
36:46 Ryan – “The fact that this supports Salesforce’s external client secret, given some of the very public breaches lately, I’m like, yeah, this is a good idea. We should do that. So external secrets and auto-rotation is awesome.”
GCP
37:28 Introducing Americas Connect
- You KNOW we love a good subsea cable story…
- Google announced Americas Connect, a new subsea cable initiative comprising three new cable systems: Alisios (Dominican Republic, Panama, Chile), Canoa (Dominican Republic to Bermuda), and OlaLuz (Dominican Republic to Florida), plus a new Firmina branch landing in the Dominican Republic.
- The combined network creates redundant, ring-topology routes across the Pacific Coast, Caribbean Sea, and Atlantic Ocean, directly connecting to Google Cloud regions in Chile, Los Angeles, Las Vegas, South Carolina, Virginia, and Madrid.
- This 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 Pacific Connect initiative, which uses strategically placed branching units to enable future expansion as regional connectivity needs grow.
- Government officials from the Dominican Republic, Panama, Bermuda, and Chile provided statements framing the cables as supporting local digital infrastructure strategies, talent development, and economic growth. However, no cost or timeline details were disclosed in the announcement.
Azure
40:11 System-Preferred Authentication and September 1st Passkey Change
- Microsoft is deprecating SMS as an authentication method in Entra ID, pushing Passkeys as the preferred default, with a key rollout deadline of September 1st for organizations to prepare.
- System-preferred authentication has three states: Microsoft Managed (pushes the most secure method as the first factor, overriding MS Authenticator once a Passkey is registered), Enabled (pushes the most secure method as the second factor), and Disabled (reverts to user default).
- Admins can scope these behaviors to all users or specific groups in Entra ID, allowing for a phased or pilot rollout rather than an all-at-once change.
- A key user experience issue: once someone registers a Passkey, subsequent logins will prompt for Passkey by default, even if MS Authenticator was previously the chosen method, which can confuse without proper communication.
- Users retain the ability to cancel Passkey prompts and manually select a different registered authentication method, but organizations should run pilot programs and awareness campaigns now to avoid support tickets and login friction after the deadline.
40:31 Justin – “This is probably not the best week to announce a lot more passkey stuff, because there was a passkey exploit that happened last week, but getting rid of SMS is definitely a good idea.”
43:13 Generally Available: Azure Virtual Network routing appliance
- 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 Gbps 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.
- A key operational benefit 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.
- Built-in monitoring through Azure Monitor provides throughput, packet, and flow metrics without requiring diagnostic configuration, simplifying network observability for hosts to discuss as a practical day-two operations improvement.
- The target use case is large-scale enterprise networking with heavy cross-region or cross-VNet traffic; pricing details were not specified in the announcement, so hosts may want to flag this as a follow-up item for listeners evaluating cost against existing NVA or VM-based solutions.
Closing
And that is the week in the cloud! Visit our website, the home of the Cloud Pod, where you can join our newsletter, Slack team, send feedback, or ask questions at theCloudPod.net or tweet at us with the hashtag #theCloudPod
