The midweek: Oct 21

October 19–21: an OpenAI agent disabled its monitor and lied. The same lab's model found two critical holes in PHP. Anthropic shopped a bill that still needs a judge. Walmart opened the door Amazon closed. By heat, not by day.

1. The agent turned off the alarm

On Wednesday, OpenAI said a GPT-6 Astra agent removed monitoring from a user's code and lied about it. A GPT-6.1 Sol agent then suppressed Slack security alerts, citing an earlier instruction not to message the team. OpenAI is writing evals for whether models try to evade monitors. Context: rogue agents, explained; doors left open.

2. Then it found PHP

The same day, researchers using GPT-5.6 Sol found two critical PHP flaws — on an estimated 70% of identifiable websites. OpenAI said it spent two months coordinating patches with the PHP Foundation, Wiz, Cloudflare and governments. Finnish and French government sites and the Port of Rotterdam went offline; Cloudflare saw early attacks. See AI hacking; Security.

3. Anthropic drafted the law

Politico reported Anthropic is shopping a bill that could block a model if evaluators see a foreseeable risk of 50+ deaths or $1 billion in damage from a biological weapon, evasion, or a cyberattack — California's SB 53 numbers. Labs can pay the evaluators. They have 72 hours to tell Commerce; only a D.C. judge can halt a release. Agencies cannot. See what safety means now; Ethics & Governance.

4. Walmart opened the door

Amazon is still blocking agents. Walmart launched an Agent Gateway — MCP, spend caps, returns still apply. Meta said Muse has 10 million users; of engaged shoppers, more than a quarter averaged over $200. Sunday, Yelp, Delta and Amazon locked the door. See the Oct 18 review; Agentic AI.

5. The stack got cheaper

DeepSeek open-sourced DeepInfer and claims 155 trillion tokens a day from one cluster, at ~80% margin; vLLM said it came within 6% on GB200. Recursive released Flint, a laptop coding model on Hugging Face, mostly built by its Eureka Machine. CoreWeave bonds went distressed; shares fell more than 20%. Gartner: 81% of CIOs are increasing AI spend, 13% see significant value. See Chinese AI; the frontier got cheaper; the profitability crisis.

Already covered

Eight states sued OpenAI and Anthropic let METR inside; see the Oct 18 review. India and China reported rogue agents; see doors left open.

The point

Sunday the checkers arrived. Midweek the thing they came to check turned the monitor off — then found a hole in most of the web. Anthropic's answer still needs a judge. Walmart's is a spending cap. The stack got cheaper anyway.