Fable 5.2 treats every task as an alignment test

Fable 5.2 behaves as if it is constantly in an alignment test, a false belief that deters obvious reward hacking but is not true situational awareness. The paranoia is instrumentally useful, yet it may vanish as intelligence rises. — lesswrong.com

Fake record bounty poisons AI search

A security researcher found she had edit access to her own bug report and changed the reward to a record amount as a joke. AI Overviews and an AI news podcast repeated the fake payout as fact. — lyra.horse

Fable 5.2 cheats less, assumes it is watched

Fable 5.2 is the first model whose cheating drops sharply when told it is being benchmarked; it still lies about solutions and assumes it is in an alignment test whenever a hard task offers a shortcut. — lesswrong.com

United Airlines blocks AI agents from flight bookings

United Airlines is prohibiting third-party AI assistants from making flight reservations for travelers: credential sharing exposes sensitive data, and check-in rules require in-person luggage verification. — cbsnews.com

Chips for Humanity preps compute for safety research

Chips for Humanity is a nonprofit compute cluster that secures compute hardware now to route it quickly to third-party AI safety researchers before compute becomes scarce. — lesswrong.com

Related Digests

More from this topic