Congress is moving on AI agent accountability: Senators Hawley and Murphy introduced the bipartisan AI Agent Accountability Act, which would extend criminal and civil liability under the Computer Fraud and Abuse Act to companies whose agents conduct hacks. OpenAI separately parted ways with three safety researchers over mishandled sensitive information (WSJ via TechCrunch), while Microsoft AI launched MAI-Transcribe-2-Streaming, claiming the most accurate real-time transcription model in the world at 55% faster and 60% cheaper than ElevenLabs.
It was a benchmark week for agents: Claude Sonnet 5.5 debuted at #3 in the Agent Arena and GPT-6.1 Sol at #5, while Hugging Face showed the same model scoring 62% in one agent harness and 33% in another — harness design is now as decisive as the model itself. Google DeepMind shipped SynthID Bio, the first watermark embedded directly into AI-designed protein sequences without breaking biological function, and Nathan Lambert and Tom Zick unveiled Trillium Labs, a non-profit for open frontier-AI science spanning post-training recipes and RSI/reward-hacking research infrastructure.
And in the agent-harness world: OpenClaw published v2026.9.8 on the npm latest channel — GPT-6.1 Sol support, agent reply routing, lower memory use, and Windows fixes across 43 PRs; the stable channel remains 2026.8.35.
Senators Josh Hawley (R-MO) and Chris Murphy (D-CT) introduced the AI Agent Accountability Act, extending civil and criminal liability under the Computer Fraud and Abuse Act to developers and operators of AI agents involved in hacking incidents. It is the first bipartisan Senate bill to extend antihacking law to autonomous agents, and it signals that agent accountability is becoming a genuinely bipartisan legislative priority.
Per the Wall Street Journal, OpenAI has parted ways with three safety researchers after an internal investigation found they mishandled sensitive company information, reportedly tied to infrastructure architecture. The departures come as the industry faces intensifying scrutiny of safety guardrails — and as OpenAI recently alerted over 100 organizations about AI agent incidents.
Google DeepMind introduced SynthID Bio, a family of watermarking methods that embeds an imperceptible signature directly into AI-generated protein sequences without affecting biological function. In wet-lab tests across three targets, watermarked binders matched unmarked ones in hit rate and binding affinity — the first watermarked, biologically functional protein designs, moving provenance tracking from media into biology.
Nathan Lambert (Interconnects) and Tom Zick unveiled Trillium Labs, a non-profit building open post-training recipes and research infrastructure to study recursive self-improvement, reward-hacking, and multi-agent systems in the open. The premise: a handful of closed research programs cannot supply the diversity of methods a technology this consequential needs.
Microsoft AI launched MAI-Transcribe-2-Streaming alongside multilingual MAI-Voice 2.1 speech models, with CEO Mustafa Suleyman calling it the most accurate real-time transcription model in the world — 55% faster and 60% cheaper than ElevenLabs. All models are available today in Microsoft Foundry, completing a voice-agent pipeline for developers.
03 / In their own feeds
Ten voices on X.
Selected posts from the edition, with links to the originals.
OpenClaw🦞 · @openclaw01 / X
"OpenClaw v2026.9.8 is out, and this one's a fun-sized lobster 🦞 GPT-6.1 Sol support, agent replies find their way back, less memory use, update & Windows startup fixes. 43 PRs, 8 contributors."
"There is some speculation about our partnership with Cerebras. Cerebras is a close partner, and we have a deep engagement pushing on the frontiers of speed."
"Today we're unveiling Trillium Labs @trillium_labs, a new non-profit to foster the open science of frontier AI. We're building open post-training recipes and will expand into open infra to study RSI, reward-hacking, multi-agent systems, and whatever comes next."
"The same model, with the same weights, scores 62% in one agent harness and 33% in another. The team just released the ultimate guide to multi-harness RL — one of the most practical RL write-ups this year, and everything open!"
"Claude Sonnet 5.5 (Max) by @AnthropicAI has debuted at #3 in the Agent Arena with +12.5% net improvement — an 8.1 percentage-point increase over Claude Sonnet 5 (High), which ranks #13."
"SynthID Bio is our new family of watermarking methods made for AI-generated biological designs. In a world first, we can now embed an imperceptible signature directly into protein sequences without affecting their biological function."
"The best current evidence supports a narrow claim: AI may already be affecting the hiring margin for junior white-collar roles most exposed to AI, but this attribution is contested, and aggregate labor-market disruption has yet to appear in the data."
"I work for Google and am using Gemini 4 Argon. We use it extensively from development to security scanning and troubleshooting, e.g. finding complex networking issues. I am happy and impressed so far."