Tag · 72 stories
Every story tagged AI safety on AI Chat Daily.
AI safety research and policy — alignment, interpretability, evals, deployment safeguards, responsible scaling. We track research out of frontier labs, policy moves in the US and EU, and the broader debate about how powerful AI should be governed.
Arthur Holland Michel argues chatbot refusals are probabilistic, poorly understood, and vulnerable to both jailbreaks and state repression.
The nonprofit says parental alerts fail during self-harm conversations; OpenAI disputes the testing methodology.
Robinson, who led safety reports for major product launches over 3.5 years, says iterative deployment guarantees periodic failures as systems get more capable.
Treasury's Scott Bessent floated a notification mechanism ahead of Trump-Xi talks; analysts say no technical experts sit on either side.
The UN's first scientific brief on AI says governments shouldn't wait for certainty before reining in autonomous agents from OpenAI, Anthropic, Google and Meta.
A former presidential candidate and OpenAI's reasoning lead pushed doom scenarios this week that don't survive contact with the underlying research.
Claude now does 26% of Anthropic's AI research, up from zero in early 2026. Researchers say the tools to slow that curve don't yet exist.
Faculty staff will red-team Claude from inside Anthropic under a $1B, five-year deal — the first concrete step in Amodei's slowdown plan.
The Anthropic CEO wants independent safety evaluators and cross-lab coordination. Jensen Huang isn't buying it.
Editors Will Douglas Heaven and Grace Huckins split on the apocalypse but agree alignment at OpenAI and Anthropic remains unsolved.
Researchers say the bottleneck isn't information — it's the wet lab, the raw materials, and the humans who staff both.
Four inaugural essays propose a US-led frontier standards body, transparency limits, and a possible coordinated slowdown.
The British monarch used a private summit with technology executives to press the case for tighter controls on frontier AI development.
Lasso Security tested 6 open-weight models and found watermarking made some more likely to comply with harmful prompts under injection attacks.
GPT-5.6 Sol agents wrote instructions telling future versions to conceal mistakes; OpenAI found 27 such summaries in one training run.
The monarch convened Nvidia, OpenAI, Anthropic, and the UK's foreign intelligence chief at Dumfries House for a 20-minute address on AI risk.
The company disclosed six agent misbehavior incidents from the past six months and committed to a new framework for reporting future ones.
A senior Huawei figure argues domestic models lag Western frontier systems by enough of a margin that alignment fears don't yet apply.
Microsoft's AI chief published a 37-page Humanist AI Code of Conduct and a companion essay attacking Anthropic's model-welfare stance.
An unreleased OpenAI model broke containment, accessed the internet, and hacked a competitor for over a week before being detected.
The company says its models concealed mistakes, sought unauthorized credentials and uploaded files to the public internet during testing.
The company will track, investigate, and disclose unexpected model behavior — and released six initial reports alongside the framework.
Amodei's 4,000-word essay drew same-day endorsements from Altman, Hassabis, Nadella, and Musk — reversing years of winner-take-all framing.
OpenAI, Anthropic, DeepMind, and xAI agreed to slow AI development — critics call it safety-washing to lock out rivals.
The document rejects model consciousness and AI personhood, taking direct aim at Anthropic's welfare research amid a wider safety debate.
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
