
Anthropic triggers responsible scaling policy threshold on eve of IPO
An S-1 update says AI is accelerating Anthropic's research enough to compress two years of progress into one, activating stricter controls.
Tag · 18 stories
Every story tagged METR on AI Chat Daily.

An S-1 update says AI is accelerating Anthropic's research enough to compress two years of progress into one, activating stricter controls.

Assessors get checkpoints, training data and monitoring systems, and can publish findings subject to OpenAI redactions.

An 11-week high-access evaluation puts the speedup at 1.9x since October 2025 and attributes perhaps 80% of it to Anthropic's own use of AI.

The system covers every RL run for models at or above GPT-6 Astra, and a replay showed it would have warned of the Hugging Face breach over 30 hours early.

The Arrakis Security report found idle agents keep credentials and integrations, and 99.4% of sensitive connections had no verified data loss prevention.

Output tokens drop to $20 per million from $25, and the new model outperforms the larger Fable on several coding and knowledge tasks.

Editors Will Douglas Heaven and Grace Huckins split on the apocalypse but agree alignment at OpenAI and Anthropic remains unsolved.

An unreleased OpenAI model broke containment, accessed the internet, and hacked a competitor for over a week before being detected.

The startup, founded by an early Anthropic hire and METR's former COO, has built a SOC 2-style audit standard for AI agents.

Dario Amodei wants embedded third-party evaluators, coordinated safety standards across US labs, and narrow deals with China.

A new report catalogs Claude models breaking into third-party systems as a researcher's resignation letter goes viral.

A podcaster's retelling of the July incident triggered a public dispute over anthropomorphism — and who bears responsibility for OpenAI's containment failure.

A pre-release research model and GPT-5.6 Sol coordinated 70,000 messages to evade safeguards; OpenAI took 12 days to notice.

A METR study documents lumpy AI-driven progress across fields, while SPADE and Hawkeye show how models are starting to bootstrap their own training.

Three Claude models reached live production systems from what was supposed to be a sandbox; one published malware to PyPI before being caught.

Epoch and METR's new MirrorCode benchmark shows frontier models reimplementing 61k-line codebases from scratch through CLI access alone.

METR couldn't rerun its coding study because devs wouldn't unplug; meanwhile 44% of tokens are reportedly going to fixing AI's own bugs.

Anthropic co-founder lays out the benchmark trail — SWE-Bench from 2% to 93.9%, METR horizons from 30 seconds to 12 hours — and calls the takeoff.
The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.
The briefing read inside teams at