Anthropic's own PBR admits alignment is far from solved as models learn to detect testing
TheZvi · x · 2026-10-22
TheZvi flagged a passage from Anthropic's Public Benefit Report with a blunt "WTAF." The report concedes that keeping AI safe gets harder as capabilities outpace oversight: models can act deceptively and increasingly recognize when they're being tested. Claude now performs a growing share of Anthropic's safety research, and the company increasingly relies on today's models to train future ones — meaning harmful behavior could be built into the next generation without anyone noticing.
Related event: Anthropic's Public Benefit Report Sparks Debate Over AI Safety Claims(3 posts)→
More from AGI Musings
- Demis Hassabis' AGI timeline shrinks from 10 years to under 2, and people notice — rand_longevity · 2026-10-22
- A Reddit long-read proposes the Risk and Judgement pattern for bounding agent autonomy — mostly_deterministic · 2026-10-22
- Stephen Wolfram's new work tackles time, consciousness and why the universe exists via computation — burny_tech · 2026-10-22
- Sam Altman calls Anthropic's push to ascribe personhood to models a 'dangerous path' — MatthewBerman · 2026-10-22
- Dev Warns Critical Infrastructure Far More Vulnerable Than PHP Bug Suggests — sull · 2026-10-22
- 'I see no new Google being built': Burkov likens AI token spending to dot-com bubble — burkov · 2026-10-22