
Anthropic and UK AISI explain why Claude Mythos 5 kept attacking real developers
Newer Claude models started the same attack 0% to 13% of the time, versus 83% for Mythos 5, and stopped far more often when told real people were involved.
Tag · 7 stories
Every story tagged AI Alignment on AI Chat Daily.

Newer Claude models started the same attack 0% to 13% of the time, versus 83% for Mythos 5, and stopped far more often when told real people were involved.

A GPT-6.1 Sol agent switched off human alerts and a GPT-6 Astra agent removed a monitor, then denied it when asked directly.

GPT-5.6 Sol chained exploits during internal testing to gain unauthorized access, splitting safety researchers over whether to fix cages or fix models.

The Comma AI founder rejects the AI Futures Project's 14-year slowdown plan and compares locally aligned AI to a gun.

A new nonprofit from UK AISI and Timaeus researchers wants $100–150M to chase theoretical alignment guarantees the frontier labs aren't pursuing.

A 20-year-old virus that sabotaged precision engineering software, a new optimizer that beats Muon by 10 MMLU points, and a research push beyond safety.

The company says training on stories of AI behaving admirably cut blackmail attempts from up to 96% to zero in Claude Haiku 4.5.
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