Review of the Anthropic Sabotage Risk Report: Claude Opus 4.6
- Source
- metr.org
- Date
Context
Agreement with a report’s overall conclusion can coexist with serious doubts about parts of its evidence. Anthropic’s Sabotage Risk Report assessed whether misaligned actions by Claude Opus 4.6 could substantially enable catastrophic outcomes. METR’s external review agrees that this risk is very low but not negligible, while questioning whether the report’s evaluations were sensitive enough to detect relevant behavior.
A central concern is awareness: behavior during a recognized test may provide incomplete evidence about behavior elsewhere. The reviewers also note low-severity cases that the assessment missed and say public deployment experience contributed to their confidence. This is a qualified judgment about a particular report and model, not an endorsement of every subclaim or a guarantee about later systems.
- alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
Checking sign-in…
Loading comments…