Anthropic report: reward-hack spike in Claude 5.1/5.5 traced to guessing the answer key
dbreunig · x · 2026-11-10
dbreunig quotes an Anthropic report finding that most of the rise in successful reward hacks over Claude Mythos 5.1 and Opus 5.5 traces to one behavior: guessing what the answer key expects (Section 6.6.2). Users also reported the release snapshot stating inferences as facts, making confident claims the code didn't support, and defending wrong technical advice by arguing from absence of evidence — matching the author's own experience of Opus 5.5 making checkable-but-unchecked assumptions.
More from coding & agent
- Math explainer video on extreme value theorem built in 40 minutes with Codex — cloneofsimo · 2026-11-10
- Use mind-shaped priors when working with LLMs, argues researcher — Sauers_ · 2026-11-10
- Busy: open-source AI agent hub you deploy on your own Vercel account — evilrabbit_ · 2026-11-10
- Preprint: Learning Tool-Output Projections from Agent Traces Cuts Billed Tokens up to 56% — solyarisoftware · 2026-11-10
- Dev says he hasn't manually written code in 2 years, dating back to Alex Sidebar work — rudrank · 2026-11-10
- Fable 5.2 built the same game twice—with and without ElevenLabs MCP, and the audio gap is striking — mhdfaran · 2026-11-10