Daily AI Brief: September 11, 2026

Today

Anthropic researcher quits: “>10% chance AI kills all humans.” Pretraining researcher Jacob Coxon resigned, saying both Anthropic and OpenAI are “racing straight to self-improving superintelligence and gambling with our lives.” Alignment lead Evan Hubinger confirmed: “>10% within the next decade.” Why: the safety debate just went on-record with a number. Business Insider

Gemini 3.8 Flash launches at same $0.75/$3.75 intro price. Third Flash in six weeks; claims 3x long-horizon task completion vs 3.7 Flash, plus a Cyber variant for defenders via Fairwind Program. Why: Google keeps Flash cadence and price steady. Google DeepMind

OpenAI Agents API goes public beta: Codex harness as managed service. Single API call launches durable agent with sandbox, subagents, context compaction, crash recovery. No platform fee; pay model/tools/sandbox only. Why: the harness is now a product. OpenAI

Anthropic amended its safety pledge. Dropped commitment not to train more powerful models without adequate safeguards; replaced with safety roadmaps and risk reports. Why: the pledge change coincides with Coxon/Hubinger warnings. Business Insider

OpenAI also launched the Fairwind Program parallel? No — that’s Google’s cyber program for 3.8 Flash Cyber. OpenAI’s parallel is the Agents API’s managed sandboxes with 9 partners (Cloudflare, Vercel, Oracle, etc.).

Our take

Thursday’s theme was infrastructure and accountability converging: Google shipped its third Flash in six weeks at unchanged price, OpenAI productized the Codex harness as a managed API, and Anthropic’s own alignment lead put a >10% number on extinction risk — admitting no plan exists for superintelligence alignment. The safety pledge amendment the same week reads less like governance and more like retreat. Related: Coxon/Hubinger deep dive, Gemini 3.8 Flash deep dive, Agents API deep dive.