Codex Kills Approval Policy, vLLM Ships Breaking KV Cache

Tags
digest
vllm
agent-safety
inference
anthropic
AI summary
Published
November 17, 2026
Author
cuong.day Smart Digest
โšก
TLDR: The safety boundary around AI coding agents just got thinner - Codex retired its `untrusted` approval policy without deprecation, and enterprise teams are furious. Meanwhile vLLM v0.34.0 shipped a breaking change to memory budgeting with extensible KV cache on by default, and Anthropic dropped $100M to mint 10,000 'Frontier Deployed Engineers' by 2027. The theme: the infrastructure is maturing faster than the guardrails.
Today is one of those days where the *plumbing* news is actually the story. Between vLLM v0.34.0 (1,061 commits, 412 contributors), SGLang 1.0 hardening its config surface, and llama.cpp pushing ternary quantization to ARM NEON, the inference stack is professionalizing at speed. But on the agent side, Codex quietly removed a safety control while OpenClaw fights 15+ P1 bugs and a CVSS 8.6 token leak. If you run agents in production, read the next section twice.

Why did Codex retire the untrusted approval policy - and should you care?

๐Ÿšจ
Breaking change with no migration path: Codex removed `approval_policy="untrusted"` without a deprecation window. Enterprise security teams relying on it as an execution boundary now have to rebuild their policy layer from scratch.
Here's the thing: approval policies are the *last mile* of agent safety. They decide whether a model can run a shell command, write a file, or touch your credentials. Removing one - especially one named untrusted, which by definition gated the riskiest operations - without a deprecation cycle is a breaking change that reads as a priority signal: velocity over safety. The pushback from enterprise adopters has been sharp, and rightly so.
This isn't happening in a vacuum. The ecosystem is simultaneously converging on zero-trust execution models - per-agent dev runners, scoped credentials, plugin identity management - as the *correct* answer to agent security. ZeroClaw is building exactly this (scoped credentials, cost enforcement, model registry standardization) and is now judged the strongest near-term deployment bet alongside Hermes Agent. The Secret-Broker Pattern is emerging as the credential-access answer: a broker mediates secrets instead of handing them to agents. And lend proposes sandboxed execution that never exposes credentials at all.
Meanwhile the threats keep stacking up. OpenClaw issue #65624 flags a CVSS 8.6 Mattermost token exposure - slash commands leaking reusable tokens via cleartext URLs. Windows 11 is reportedly telemetry-sniffing which coding agents you use, drawing Microsoft sharp criticism. And Anthropic just shortened its Coordinated Vulnerability Disclosure windows from 90 to 30 days for high/critical issues (45 for others) effective January 1, 2027, explicitly because AI-accelerated vulnerability discovery makes the old timelines unsafe.
  • Codex alpha builds (rust-v0.171.0-alpha.8-11) shipped app-server thread resume refactor, SSH HTTP/2 fallback, and undo/redo in the prompt composer - but no stable release.
  • Claude Code v2.1.326 added Code Mode support plus startup validation warnings for malformed `--allowedTools` configs - a *good* safety counter-example.
  • GitHub Copilot CLI v1.0.113-1 now enforces managed policy via `features.browserUse`.
  • Kiro CLI is being cited as the reference for building a control layer that bounds uncontrolled code generation.
Retiring a safety control without deprecation isn't a breaking change - it's a values statement. Enterprise teams should treat it as one.

Which inference engines shipped breaking changes today?

๐Ÿง 
vLLM v0.34.0 is the biggest inference release of the day: 1,061 commits, 412 contributors, 10 new models, and extensible KV cache enabled by default. Your memory budgeting math just changed.
Extensible KV Cache dynamically sizes the cache post-warmup using free memory, which is great for long-context efficiency but means static memory budgets you tuned last month may now behave differently. If you run vLLM in production, re-benchmark. Alongside it, the MoE Expert Offloading RFC (#38256) proposes an async pipeline + GPU cache for pushing experts beyond VRAM limits - a big deal for serving giant MoE models on constrained hardware.
SGLang 1.0 is the other headline: the release tracker (#50045) is active, silent config changes now become *errors*, and breaking changes from 0.5.x are expected. The SGLang Rust Server is migrating its final control-API routes (PRs #49742-#49748), cutting Python overhead. Just watch HiCache + Mooncake - documented cache hit-rate dips under L2/L3 capacity pressure are a real problem for long-context serving at scale.
On the model side, the floodgates opened. vLLM added native support for Mistral Large 4, Kimi K2.5/K3 Flash (AMD AITER FlashKDA on CDNA4), Qwen3.8-2.4T-A95B / Flash-Next (full ROCm), MiniMax-M3.1-Flash, and Nemotron-3.5-Transcribe. DeepSeek-V4.1-Flash got its SM120 CUDA graph/decode throughput fixed (PR #65527) but *still crashes on SGLang*. GLM-5.3-Flash runs on Huawei Ascend A3/A5 NPUs in SGLang and via CANN/OpenVINO in Ollama.
  • llama.cpp builds b12453-b12441: RDNA3_5, gfx1013, WebGPU per-head flash attention, Android Vulkan, modular UI refactor (`/ui/apps/chat`), plus MiniCPM5 MoE (gated GQA) and Kokoro TTS / tiny-voca for local voice agents.
  • Experimental ternary quantization: PTQ1_0 (group 128) and STQ1_0 (sparse ternary for ARM NEON) target edge memory reduction.
  • Ollama: VRAM accounting/GPU splitting fix (#19632), MLX speculative decoding, Huawei CANN backend (PR #19732), per-request metrics via OLLAMA_LOG_METRICS (PR #19723), and PPLX Decider v1.3 support.
  • Unsloth v0.1.918-beta blends LoRA fine-tuning, local serving, CPU embedding serving, and MCP-based computer-use.
  • LiteLLM v1.108.10 adds cosign verification to its multi-provider gateway.
  • Avoid W4A16 - the HummingLinearKernel produces non-deterministic outputs across processes (vLLM #66232).
โš ๏ธ
Hardware specialization is now mainstream: AMD CDNA4/gfx950 optimization campaigns span vLLM (MXFP4 dequant, FlashKDA), llama.cpp (GDN MMA), and Unsloth. Meanwhile Claude Mythos claims it autonomously wrote nearly 100% of kernels, cutting AMD GPU onboarding from up to 12 months to ~3 days - a direct threat to NVIDIA CUDA's moat.

Is the agent 'plugin economy' replacing general-purpose models?

The clearest trend across GitHub today: Agent Skills are eating the ecosystem. Instead of one model to rule them all, we're getting modular, evaluable, reusable domain-specific capabilities - and the star counts prove it. NVIDIA/SRL Evaluator (SkillEvaluator) formalizes this with quality gates, synthetic datasets, and live behavior measurement. MCP has become the connective tissue, showing up across trending repos as the de facto agent-tool communication standard.
  • google-labs-code/stitch-skills - a standardized skills library compatible with Claude Code, Gemini CLI, and Cursor. Cross-ecosystem interoperability is finally real.
  • md2video-audio - Markdown to MP4 with AI voiceovers via Marp. Top community interest in the Claude Code skills queue.
  • proofcore-contract-auditor - static analysis of Solidity/Rust contracts with cryptographic audit proofs anchored on TON blockchain.
  • Claude-BugHunter - 82 bug-hunting skills + 15 slash commands across 24 vulnerability classes.
  • kajisho5/ffmpeg-skill - 42 FFmpeg tools, zero cloud dependencies, for privacy-first video editing.
  • claude-trading-skills, daily_stock_analysis, HKUDS/Vibe-Trading - a full trading-agent stack is forming.
  • chrome-devtools-mcp - Chrome DevTools reimagined as an MCP server for browser automation and debugging.
  • alibaba/open-code-review - hybrid deterministic + LLM code review with 10+ language security rulesets battle-tested at Alibaba.
Memory is the other half of the plugin story. claude-mem, Mem0, and Volcengine's OpenViking (unifying memory + RAG + skills) are all racing to be the persistent context layer. Graphify turns codebases into queryable knowledge graphs with no vector store; RAGFlow has evolved into full context infrastructure. Even ChatGPT Memory is in the news for the wrong reason - reports of confident false-memory reconstruction due to relational context loss.
๐Ÿ 
Local-first is the new default posture: Rakazo (open-source Grok alternative, 401 votes on Product Hunt) leads with user ownership and self-hosting. ECC (agent harness for Claude Code/Opencode/Cursor), Univer, OpenViking, and Ollama's NPU expansion all point the same direction - away from vendor lock-in.

Are agents becoming coworkers - and can the frameworks handle it?

The framing shift is unmistakable: agents are evolving from code generators into autonomous coworkers with remote control, peer-to-peer messaging, and task coordination. Taskhold gives teams a shared task list with live progress across Codex clients. TeapotRouter solves key fragmentation with one metered API key for both Codex CLI and Claude Code. Talyn runs Claude Code and Codex in parallel cloud sandboxes. Pixel Office literally gamifies managing a team of agents.
But the frameworks keeping the lights on are in stabilization mode. OpenClaw is the most active project in the ecosystem (318 issues / 500 PRs in 24h) but carries 15+ open P1/P2 bugs around session state corruption, memory leaks, and platform crashes. Prerelease betas v2026.11.5-beta.1-4 bring historical-state recovery fixes (#164445) and better agent yield handling - no formal release notes yet. The long-requested one-way dispatch mode (#44309), enabling agent-to-agent handoffs without ping-pong, is predicted for v2026.11.6.
  • Hermes Agent v0.23.2 (Nous Research) - rated healthiest at 4.2/5, rolling up ~1,989 merged PRs since v0.23.1 across GitHub CLI, Microsoft Store, Docker, and Hermes Cloud. No breaking changes; full changelog deferred to v0.24. New /implement command (PR #156782) runs saved plans.
  • QwenPaw (AgentScope) - multi-channel WeChat/DingTalk agent framework, 46 issues / 50 PRs, v2.3.1 hotfix incoming. Plugin SDK 1.1 (PR #8764) adds plugin identity, permissions, and lifecycle management.
  • ZeroClaw - security-first orchestration, driving model registry standardization and cost enforcement.
  • IronClaw (NEAR AI) - zero activity in 24h, rated 2.0/5. Likely abandoned.
  • Pi - no release but leads technical maturity with 10 PRs fixing race conditions, CPU bloat, and stream leaks on an event-driven kernel.
  • Qwen Code v0.27.4-nightly - headless event-bus kernel and unified agent loop.
  • OpenCode - `@opencode/pty@0.6.0` for daemon lifecycle + PTY protocol stability, plus community heat over free-tier quota transparency.
Session continuity is now table stakes - everyone's converging on auto-recovery after network loss, usage limits, and WebSocket timeouts. And the reliability consensus is clear: session integrity, cost transparency, resource limits, and secure isolation now outrank feature velocity as adoption drivers. Idempotent retry patterns are becoming mandatory to prevent duplicate charges from tool-call retries. Opus 5 users are hitting a regression where prose between tool calls gets summarized into 'narration' blocks and code blocks get dropped - breaking debugging flow.

๐Ÿ“Š CLI Agent Tooling: Where Things Stand

๐Ÿ“Š Tool | What Shipped | Why It Matters

  • **Claude Code** โ€” v2.1.326 - Code Mode, agent teams, `--allowedTools` validation โ€” Safety validation is the right pattern; skills ecosystem booming
  • **OpenAI Codex** โ€” rust-v0.171.0-alpha.8-11 - thread resume, SSH HTTP/2, undo/redo โ€” Alpha only, and `untrusted` policy retired without deprecation
  • **GitHub Copilot CLI** โ€” v1.0.113-1 - managed policy via `features.browserUse` โ€” Mature cadence, enterprise policy enforcement
  • **Gemini CLI** โ€” No release - POSIX-native execution, AST-aware context reduction โ€” Community-driven optimization focus
  • **OpenCode** โ€” `@opencode/pty@0.6.0` - daemon lifecycle, PTY stability โ€” Free-tier quota transparency pressure mounting
  • **Pi** โ€” No release - 10 PRs on race conditions, CPU, stream leaks โ€” Technical maturity leader via event-driven kernel
  • **Qwen Code** โ€” v0.27.4-nightly - headless event-bus kernel โ€” Scalable orchestration via unified agent loop

โšก Quick Bites

  • Claude Frontier Academy - Anthropic's $100M initiative to train 10,000 enterprise 'Frontier Deployed Engineers' by 2027. Launch cohorts: Accenture, Bain, Deloitte, Morgan Stanley, Novo Nordisk, Commonwealth Bank of Australia.
  • Anthropic's Public Benefit Fund - committed shares (5% initial, +2% per 50% valuation increase) to an automatically growing philanthropy and labor-disruption fund. Christopher Stubbs (Harvard physicist, co-discoverer of dark energy) joined the Long-Term Benefit Trust.
  • Anthropic - $10M CAD to Canadian AI research, new Singapore office with government MOU, and a $5M wellbeing research grant program for independent AI impact studies.
  • Barclays - scaling Claude enterprise-wide across software dev, legacy modernization, and customer service in a regulated banking environment.
  • Qwen4-Max - leap in reasoning and code generation, but sparking debate over alignment and transparency.
  • OpenAI - eight metadata-only index pages (healthcare, academics, child safety, assessments) signal pre-launch vertical products. ChatGPT for Healthcare with computer use, Embedded Assessments, and an LVMH Media Division partnership all hinted.
  • ppt-master - 77k+ stars for automated PowerPoint generation with animations, charts, and audio narration.
  • Firecrawl - 202k+ stars, web data ingestion with stealth browsing for AI agents.
  • CowAgent - lightweight self-evolving personal AI assistant with memory across multi-model, multi-channel deployment.
  • Dream-Num/univer - Office SDK for embedding agents into collaborative document editing and calculation.
  • AutoGPT - still the core reference for autonomous agentic workflow design.
  • Mem - announced it's shutting down. A reminder that consumer AI productivity tools are fragile.
  • Jane Street - its only NYC datacenter nearly drowned after a false smoke alarm triggered a full shutdown. Infra fragility is real.
  • libreagi - privacy-focused platform preserving open-source AI models against hoarding and centralization.
  • db-branch - database branching in seconds for AI agents and CI using ZFS snapshots.
  • OpenCode Browse - browser-based agent tool for code exploration; raising data privacy and cloud inference concerns.
  • ByteAsk Cloud Agents, Polylane for Security, Tulvo, Agentic Canvas, Frevana Multiplayer, Spiik, Perit Marketplace, Transformer.gg - a wave of agent-native SaaS covering Slack debugging, real-time vuln scanning, adaptive workspaces, visual workflow design, business co-pilots, local meeting notes, dataset marketplaces, and influencer automation.
  • Unix Domain Sockets (Ollama #19464) - feature request for low-latency, secure IPC. Local agent inter-process communication is coming.
  • Intel Core Ultra NPU (OpenVINO) - requested Ollama backend (#19535), still not implemented. A gap vs AMD/Huawei NPU support.
  • Rust Frontend (vLLM) - experimental via `VLLM_USE_RUST_FRONTEND=1`, feature parity roadmap #44280 pending. Python/Rust API divergence expected.

๐Ÿ”ฌ Research & Benchmarks Worth Knowing

  • SWE-CoWork - benchmark simulating real software dev workflows across 101 repositories for evaluating coding agents as digital employees.
  • GEM-MM - entropy-guided preference alignment reducing hallucination in vision-language models by prioritizing informative tokens.
  • A Matrix with Steps - dual episodic + procedural memory so agents and robots avoid past failures and learn from verified successes.
  • CECAR - cache and expert co-aware routing that makes on-device MoE LLM inference viable.
  • CapCover Filter / SAMQs - semantic approximate membership queries that reject irrelevant embeddings without full similarity search. Big for vector DB throughput.
  • KV thrashing - identified as a major inefficiency in long-output LLM serving; proactive preemption minimizes state loss.
  • Gemini 4 vs Gemma 4 - hard-math benchmarking shows high-thinking modes cost 2.6-4.8x more with minimal performance gains. Turn thinking down.
  • Query-ATLAS, D'BEAST, RoboKit (2-60x speedups), FRAAC (first French multimodal LLM for pictogram AAC), MT-RGCN-DTI, acuSimNet - spanning symbolic discovery, driving explainability, robotics, accessibility, drug-target interaction, and acupuncture point localization.
  • ML Modules for Mere Mortals and The Little Book of Generative AI Foundations - two primers worth a weekend. Transformers explained as a natural extension of associative memory.

โ“ FAQ: Today's AI News Explained

  • Q: What happened with Codex's approval policy? - OpenAI Codex retired `approval_policy="untrusted"` without a deprecation window. It was a key execution safety boundary, and enterprise teams now need to rebuild their policy layers. Alpha builds (rust-v0.171.0-alpha.8-11) continue but no stable release shipped.
  • Q: What is extensible KV cache in vLLM v0.34.0? - It's now enabled by default and dynamically sizes the KV cache post-warmup based on free memory. This improves long-context efficiency but changes memory budgeting behavior - re-benchmark production workloads.
  • Q: What is the Claude Frontier Academy? - Anthropic's $100M program to train 10,000 enterprise 'Frontier Deployed Engineers' by 2027. Initial cohorts come from Accenture, Bain, Deloitte, Morgan Stanley, Novo Nordisk, and Commonwealth Bank of Australia.
  • Q: Is SGLang 1.0 a breaking release? - Yes. Silent config changes now become errors and breaking changes from 0.5.x are expected. The Rust server is migrating final control-API routes (PRs #49742-#49748).
  • Q: What are Agent Skills and why are they trending? - Modular, evaluable, reusable domain-specific capabilities replacing general-purpose model bloat. NVIDIA/SkillEvaluator provides quality gates; google-labs-code/stitch-skills standardizes them across Claude Code, Gemini CLI, and Cursor; MCP is the communication layer.
  • Q: Which agent framework is healthiest right now? - Hermes Agent at 4.2/5 (v0.23.2, ~1,989 merged PRs). ZeroClaw is the strongest near-term deployment bet for security. OpenClaw is most active but carries 15+ P1/P2 bugs. IronClaw looks abandoned at 2.0/5.
  • Q: What is Claude Mythos and why does it matter? - Anthropic claims it autonomously wrote nearly 100% of kernels, cutting AMD GPU onboarding software work from up to 12 months to ~3 days. That's a direct threat to NVIDIA CUDA's ecosystem moat.
๐Ÿ”ฎ Editor's Take: We're watching the agent stack split in two - inference infrastructure professionalizing at breakneck speed (vLLM, SGLang 1.0, ternary quant) while agent safety gets *loosened* (RIP `untrusted` policy). The frameworks that win 2027 won't be the smartest - they'll be the ones that survive a CVSS 8.6 audit. Bet on ZeroClaw and Hermes Agent; keep OpenClaw on a short leash.