Friday, October 23, 2026
TL;DR
Claude Code's 2.1.309 closed a real sandbox escape - upgrade if you run auto mode. Anthropic handed its Agent Skills standard to neutral governance, the same move it made with MCP in December. AI stocks sold off sharply on debt-financed data-center worries, and a Micron union struck in Taiwan.
Act on this
- Upgrade Claude Code to 2.1.309 if you run sandboxed or auto mode - it fixed real escapes, including shell startup files running outside the sandbox. Claude Code releases #
Signals
Major commercial platforms are choosing to block AI agents outright rather than build for them. #
Amazon blocked Meta's Muse agent, JPMorgan Chase and Charles Schwab bar agents from touching customer accounts, and Delta and United restrict automated booking traffic, per TechCrunch's Oct 6 roundup; Cloudflare has blocked agents by default on ad-supported sites since September. Zvi Mowshowitz's Oct 22 weekly roundup treats the pattern as now widespread enough to flag on its own, not isolated incidents. TechCrunch, the next hurdle for AI agents · Zvi Mowshowitz, AI #191
Beyond Zvi Mowshowitz's week-ending synthesis, it was a quiet week for new signals.
News
WATCHAnthropic hands Agent Skills to neutral governance, following MCP #
Anthropic donated its Agent Skills standard to the Linux Foundation's Agentic AI Foundation on Oct 22, the same body it gave MCP to in December 2025 and that already hosts OpenAI's AGENTS.md and Block's goose. Anthropic, donating Agent Skills to the Agentic AI Foundation
WATCHAI stocks sell off worldwide on data-center debt worries #
The Nasdaq fell 2.6% and the S&P 500 1.2% on Oct 22, the S&P's worst day since July, as investors weighed high borrowing costs for data centers and chips. CoreWeave dropped 10.1%, Meta 7%, and Nvidia 4.4%. AP, via BNN Bloomberg
WATCHMicron's Taiwan union stages the island's first chip-sector strike #
The Taoyuan union walked out for 24 hours on Oct 22 after rejecting Micron's 68-month bonus offer, seeking a permanent 15% share of operating profit as AI memory demand lifts earnings. A separate Taichung union already settled. The Star, Micron union Taiwan strike
WATCHA second Google DeepMind staffer resigns over Pentagon AI work #
Research engineer Sofia Liguori resigned Oct 21, citing Google's classified Pentagon AI contract and saying she no longer believes the company's technology nets out positive for society. AI safety researcher Alex Turner quit over the same contract in July. The Intercept, DeepMind resignation
The long view
A year ago, MCP - the agent-tool protocol Anthropic introduced in November 2024 - was still entirely Anthropic's to govern; the donation to a neutral foundation, which would also take in OpenAI's AGENTS.md and Block's goose, didn't happen until December 2025. On October 22, Anthropic handed over Agent Skills too, the same move a second time, so both of its major agent-interoperability standards now sit outside its own roof. If this holds, a lab introducing a widely-adopted agent standard is expected to eventually hand it to neutral governance, not keep steering it alone indefinitely.
Also noted
- vLLM's Oct 18 release likely fixed GLM-5.3-Flash's speculative-decoding bug; the VRAM and Metal regressions remain open. vLLM v0.32.0 release #