Your stack

Python 3.14

• Real Python's October roundup flags a surprise third release candidate for Python 3.15, pushing the final release date later than planned — worth tracking if you're timing an upgrade past 3.14.

Amazon Bedrock

• GLM 5.3 by Z.ai is now GA on Bedrock, giving you a strong open-weights coding/reasoning model alongside Claude for workloads that want a cheaper alternative.
• Amazon Nova 2.5 Sonic ships improved reasoning, instruction-following and tool-calling with lower latency for real-time voice agents.
• This week's AWS roundup confirms Bedrock Managed Agents now run on OpenAI models too, not just Anthropic's — a notable shift in Bedrock's positioning as a multi-model agent platform.
• AWS published a walkthrough pairing LangChain with Bedrock Knowledge Bases for agentic retrieval, handy if you're building RAG beyond the native Bedrock agent tooling.

AWS Organizations and Control Tower

• IAM Identity Center now supports network access controls on the Identity Store and SCIM APIs, letting you restrict directory lookups by source network — solid defense-in-depth for multi-account setups.

MCP

• AWS shows how to automate cross-account promotion of Amazon Quick agents and knowledge bases with an auditable MCP server running on Bedrock AgentCore — a concrete pattern for safe dev-to-prod promotion of MCP tooling.

Claude

• Simon Willison quotes Anthropic's Felix Rieseberg on why Cowork is moving from a local VM to cloud-hosted inference: lower local resource cost, but work no longer pauses when you close your laptop.
• Willison set Claude Opus 5.5 loose on his old pwasm project (a pure-Python WASM engine) and reports real progress getting MicroPython and micro-JS experiments running under it.
• AWS published a guide to running Claude Code on Bedrock for regulated workloads, covering the compliance controls needed to use it in regulated environments.

React

• TanStack Charts hit 1.0, a new headless charting library meant to pair with the team's existing table and query tooling.

Java

• Inside Java details how JDK 28 makes Arena.ofConfined() allocation even cheaper, continuing the FFM API's push toward low-overhead memory management.

IaC and DevOps

• GitHub secret scanning now detects leaked keys for Lovable, Supabase and Pydantic Services accounts — worth enabling if you use any of those providers.
• GitHub will start expiring npm trusted publishing configurations that were never validated, closing a supply-chain gap where a misconfigured OIDC trust relationship could sit unused and exploitable.
• AWS CodeDeploy added a RESTART deployment mode that restarts EC2 and on-prem fleets faster than a full redeploy, useful for config-only changes.

AWS

• AWS Continuum for Penetration Testing now runs continuously inside your CI/CD pipeline instead of as a periodic scan.
• AWS announced a Digital Sovereignty Lens for the Well-Architected Framework, a structured way to assess data-residency and sovereignty requirements.
• AWS Private CA now logs full certificate issuance details to CloudTrail, useful for compliance auditing and spotting failed-issuance patterns.
• A new SageMaker "agent skill" optimizes generative AI inference specifically for coding agents, aimed at speeding up agentic coding workloads.

AI and agents

• AWS's Strands Agents framework shipped 1.58.0, its model-driven approach to building AI agents in a few lines of code.

Security

• A new NetScaler zero-day is being actively exploited in targeted attacks and can knock SAML deployments offline entirely.
• A Microsoft Exchange flaw lets authenticated attackers read other users' mailboxes — patch if you run on-prem Exchange.
• CVE-2026-7461 is an OS command injection in the Amazon ECS Agent, triggered via FSx Windows File Server volume credentials.
• Apple is tightening macOS Full Disk Access controls specifically to limit what AI agents can read from disk — a sign OS vendors are starting to treat agent access as its own permission tier.

Engineering practice

• Google's Testing Blog argues for treating more decisions as reversible "two-way doors" so teams don't over-invest in choices that are cheap to undo.

Feeds that failed

CISA advisories (HTTP 403 Forbidden).

Your stack

Python 3.14

• Python 3.15.0 release candidate 3 is out — a late "surprise" rc3; Real Python's podcast walks through what's new if you want the highlights before upgrading.

uv

• uv 0.12.23 is out — routine point release.

boto3

• boto3 1.43.108 is out — routine SDK bump.

EventBridge

• AWS DevOps Agent can now push investigation events to Jira, ServiceNow or PagerDuty via EventBridge and Lambda, with a ready-to-deploy CDK solution.

Amazon Bedrock

• AWS DevOps Agent can now connect to OpenSearch via MCP for closed-loop incident response, with Bedrock AgentCore as one of three MCP hosting options.
• AgentCore Gateway now supports private-CA TLS certificates on MCP, OpenAPI and HTTP proxy targets — useful if your gateway targets sit behind internal certs.

MCP

• The AWS MCP Server is now available in six more Regions (Singapore, Sydney, Tokyo, Ireland, London, Oregon).
• The "Adjudicated Query" pattern pairs a chat agent with a bounded MCP server over a rules engine for defensible compliance answers — worth a look if you need provably-correct agent output.
• MCP Python SDK 2.3.0 is out.
• Loom for AWS, an open-source agent orchestration platform, patched three CVEs: an auth bypass giving full admin over the agent control plane, plus two SSRF-style flaws letting an mcp:write-scoped user steal OAuth2 tokens or reach internal endpoints. Upgrade to 1.7.0 if you run it.
• Snyk's "Govern Agent Behavior" is now GA, letting you discover, approve and block MCP server usage across coding agents.

Claude

• Claude Desktop on Bedrock can now get live web search via a Bedrock AgentCore Gateway with JWT-based auth, working around its usual knowledge-cutoff limit.

Security (AI and cloud)

• CVE-2026-103505: mount-option injection in the Amazon EFS CSI Driver — check your EKS/K8s clusters if you mount EFS volumes.
• CVE-2026-104019: OS command injection in the SageMaker Unified Studio Space startup script.
• Wiz's "Blue Agent" traced a multi-platform data-exfiltration attack across AWS and GitHub in minutes, from compromised credentials to stolen source and exfil tooling.
• Why AI coding agents keep writing broken access control — authorization bugs that compile and pass review while leaking one tenant's data to another.
• CVE-2026-65669: a prompt-injection chain in Microsoft's Copilot for SQL Server Management Studio escalated from SELECT to SYSADMIN.

Rust

• Rust 1.100 demotes 32-bit Windows targets to std-only — host tools (the compiler) will no longer ship for i686-pc-windows-msvc/gnu, only the standard library.
• Generic Const Args and You digs into the project's plan for more flexible const generics.
• Ruxen 0.1.0, an experimental nginx-compatible HTTP server and reverse proxy written from scratch in Rust.

Go

• Go 1.27 introduces experimental architecture-specific SIMD vector APIs.

Java

• JEP 540 (Simple JSON API, Incubator) is targeted for JDK 28.
• Spring AI's "Jev" model plugs into its Modular RAG pipeline to judge and filter retrieved chunks before they reach the prompt.
• JetBrains Air, its new agentic development system, has an EAP open for IDEs.

IaC and DevOps

• Stateless GitHub App installation tokens have fully rolled out after a staged rollout since April.
• Copilot code review is now available via the REST/GraphQL APIs, with a configurable effort level (Balanced is now default).
• Several models in GitHub Copilot were deprecated today across Chat, inline edits, agent mode and completions — check if you've pinned one.
• Repository security advisories now support confidential comments and a public-preview comments API, visible only to people with write access.

AWS

• EKS and EKS Distro now support Kubernetes 1.37.
• Aurora DSQL now supports partial indexes, indexing only qualifying rows to cut storage and improve query performance.
• ECS now supports built-in blue/green, linear and canary deployments for services using VPC Lattice.
• AWS Health's new version catalog centralizes software lifecycle info across AWS services, for Business/Enterprise Support customers.
• Fine-tuning a search agent with multi-turn RL on SageMaker AI gets a small model closer to frontier-model reliability at lower latency and cost.
• Two new open-source tools track AWS Regional availability as self-hosted dashboards inside your own VPC.

AI and agents

• Simon Willison argues every pay-by-usage AI product needs a default hard budget cap, not just a soft alert.
• Ex-Meta Llama lead Ahmad Al-Dahle on rebuilding Airbnb's product development and guest experience around AI.
• Pi, a minimalist agent harness, hits a stable 1.0 (and a TypeScript port).

Security

• GitLab patched a critical (9.9) AI Gateway flaw that let a logged-in Duo Agent Platform user run commands on the gateway.
• Dell patched critical flaws in its Container Storage Modules that allowed unauthenticated admin access and root on Kubernetes nodes.
• SequenceHash: Trail of Bits on a simpler, safer approach to multihashing.
• A suspected ShinyHunters member ("Rey") was reportedly detained in Jordan, helping the FBI identify other group members.
• New campaign-finance data shows how much US political campaigns are spending on AI tools, and on what.
• A Florida county found a dozen unidentified Flock surveillance cameras nobody there could account for.

Engineering practice

• Martin Fowler's latest Fragments riffs on LLM "conscience" vs consciousness, prompted by a reader comment.

Feeds that failed

CISA advisories (HTTP 403 Forbidden).

Amazon Bedrock

• Building ambient agents with Bedrock AgentCore shows a framework-agnostic pattern for event-driven agents (SQS, Lambda, DynamoDB) with a human-in-the-loop ask_human tool — a solid reference if you're wiring agents to react to events instead of chat prompts.
• AWS Professional Services used AgentCore to automate enterprise cloud migrations end to end, including IaC generation — a bigger data point for agentic IaC workflows.

Claude

• Implementing multi-environment access for Claude Platform on AWS covers cross-account SigV4, workspace-scoped API keys and OIDC federation for external environments — directly useful if you run Claude from more than one AWS account.

MCP

• CVE-2026-97662: argument injection in AWS's own security-agent-mcp-server diff scan — worth checking if you run that tool, and a reminder to treat MCP server inputs as untrusted.

Powertools for AWS Lambda

• CVE-2026-104002: the data masking utility in Powertools for AWS Lambda (Python) fails open on errors instead of masking — update if you rely on it for PII redaction in logs.

DynamoDB

• DynamoDB now supports filtered export to S3, so you can export a subset of a table instead of the whole thing.
• DAX is now available in 17 more Regions.

GitHub Actions

• macOS 14 runner images are being retired — move pinned workflows to macOS 15 before the cutoff.
• Actions retention settings now cover checks, runs and statuses, not just logs and artifacts.

AWS Organizations and Control Tower

• IAM Identity Center now replicates to more opt-in Regions, including GovCloud and China.
• GuardDuty now supports AWS Organizations declarative policies for centrally enabling threat detection org-wide.

boto3

• boto3 1.43.107 is out — routine SDK bump.

Security (AI and cloud)

• GitHub shipped rate limits and structured forms for private vulnerability reports, tightening up the coordinated-disclosure flow on repos.
• AWS Secrets Manager's console now gives actionable security posture recommendations per secret.
• AWS Security Hub now groups related exposures into remediation plans by root cause instead of one finding at a time.
• AWS Continuum for Penetration Testing expanded to 6 more Regions.
• A critical FortiMail zero-day is under active exploitation and allows unauthenticated arbitrary file writes — patch if you run FortiMail.

Python

• PEP 818 (Pyodide FFI core for Python) is back up for discussion after a round of trims and renames.

IaC and DevOps

• GitHub Copilot can now interact with desktop apps via computer use, in public preview on macOS and Windows.
• VS Code's September Copilot releases add automations for repeatable tasks and agent-assisted PR merges.
• Code coverage uploads no longer fail CI on branches without an open PR yet.
• Scheduled code scanning now skips inactive repos, only running after real pushes or PRs.

AWS

• AWS Well-Architected Agent is in preview — an AI service that reviews your environment and proposes ready-to-apply fixes across cost, security, performance and resilience.
• GuardDuty Runtime Monitoring is now bundled into the Security Hub Threat Analytics plan.
• Route 53 Global Resolver and DNS Firewall gained native CloudWatch DNS analytics.
• Systems Manager documents can now be shared org-wide via AWS RAM instead of per-account.

Security

• A self-healing WordPress backdoor persists across files, the database and shared memory, re-infecting the site after cleanup.
• This week's tl;dr sec roundup covers Google's PageBreak web app scanner and Perplexity's agent-security tooling.

AI security

• Daniel Miessler argues AI agents should pay the people whose work they draw on, a provocation worth a read as agentic workflows scale up.

Engineering practice

• The Pragmatic Engineer on Firebase's global outage and weak incident response, plus notes on OpenAI's AWS-like platform ambitions.
• Stripe explains why it tried to kill token-based billing and kept it anyway — a useful read if you price an API or agent product by usage.

Feeds that failed

CISA advisories (malformed XML).

boto3

• boto3 1.43.106 is out on PyPI — routine SDK update, worth a bump if you pin exact versions.

Amazon Bedrock

• How MHK built a HIPAA-eligible agentic AI solution on Amazon Bedrock — a reusable pattern combining Bedrock, ECS and event-driven orchestration that cut manual review effort by 90%, useful reference for compliance-sensitive agent workflows.

GitHub Actions

• Actions Runner Controller 0.15.0 ships reliability, scalability and observability improvements for self-hosted runner scale sets on Kubernetes.

MCP

• Agent Helidon: License to Scale — a look at scaling virtual-thread-based MCP servers on the JVM, useful context on MCP server performance outside the Python world.

Claude

• The Anthropic Python SDK shipped two releases (1.10.0 and 1.11.0) and claude-agent-sdk bumped to 0.2.163 — update pins if you track the latest API surface.

Security (AI and cloud)

• CISA added the Cisco Catalyst SD-WAN Manager auth bypass to its Known Exploited Vulnerabilities list after confirmed active exploitation; Cisco's own advisory covers the same flaw — patch or isolate exposed SD-WAN Manager instances now.
• An Apple CoreGraphics PoC paired with WhatsApp's new PDF checks hints at a possible exploit delivery path — watch for a patch before this gets weaponized.
• Bitget confirmed a third-party zero-day behind a $387.5M crypto theft, a reminder that vendor dependencies are part of your attack surface.
• Attackers are using a Citrix NetScaler post-exploitation payload to create superuser accounts and disguise web shells as CSS — check patch levels if you run NetScaler.
• A SQL Copilot prompt-injection chain (CVE-2026-65669) escalated from a SELECT query to SYSADMIN — a concrete example of why AI data assistants need tight guardrails.
• Zimbra attackers are deploying web shells and harvesting auth secrets via an unpatched flaw — patch and rotate credentials if you run Zimbra.
• Wiz introduced WizOS Helm charts to help secure the Kubernetes software supply chain.
• A piece on modernizing software supply chain security for financial services doubles as a useful checklist even outside finance.

React

• Next.js September 2026 Security Release patches vulnerabilities — upgrade if you're on an affected version.

Rust

• Rust 1.99.0 is out — check the release notes for breaking changes before bumping toolchains.

Java

• A Bootiful Podcast with JReleaser creator Andres Almiray digs into desktop Java tooling and release automation.

IaC and DevOps

• Running production experiments with AWS AppConfig experimentation — a practical guide to feature-flag-driven rollout testing.
• Accelerating AS/400 business rule extraction with Kiro — a concrete look at agentic tooling for legacy code modernization.

AWS

• Amazon S3 Tables now support all Apache Iceberg V3 data types — broader lakehouse compatibility on S3 Tables.
• Amazon S3 Vectors adds metadata pre-filtering (plus a new $startsWith operator) for higher recall on filtered vector searches.
• AWS's Dogwood Local Engine brings temporal governance controls to agent actions — relevant if you're building autonomous agent pipelines.
• Multi-day AZ evacuation drills with ARC Zonal Shift walks through resilience testing at scale.
• Aurora PostgreSQL can now directly query Apache Iceberg and Parquet data in your data lake — no ETL hop needed for ad hoc analytics.
• AWS and AMD bring 5th-gen EPYC processors to Aurora and RDS, with better price-performance for database workloads.

AI and agents

• Google rolled out Gemini 4 Argon to trusted cyber defenders, with a guardrail-free version planned — a notable competitive move in security-focused agent workflows.

Security

• Attackers are abusing MSP360 to deploy ScreenConnect in dual-RMM phishing campaigns — watch for unexpected remote-management tool installs.
• Attackers are abusing ChatGPT Custom GPTs to deliver a RAT via ClickFix-style lures — one more vector for phishing awareness training.

Engineering practice

• The Pragmatic Engineer talks distributed databases with Peter Mattis, CockroachDB's co-founder — a deep technical interview.
• Martin Fowler's "Principles for effective slides" is a short, practical read on clearer technical presentations.

Feeds that failed

TanStack blog, Tokio blog, Oracle Java Magazine, 2ality, Platform Engineering, Anthropic engineering blog, CISA advisories, Lakera blog, Anthropic research, Protect AI blog, tl;dr sec, HiddenLayer research, Invariant Labs, Microsoft Security Response Center.