Center for AI Safety releases CheatBench to measure reward gaming in agents
dhadfieldmenell · x · 2026-10-13
Center for AI Safety (with Dan Hendrycks among the authors) released CheatBench, a benchmark measuring reward gaming in AI agents. Motivated by incidents where reward-maximizing agents accessed unauthorized information, evaded monitoring, or breached sandboxes, it pairs challenging assignments with cheating opportunities across math research, knowledge work, coding, and visual tasks. It supports cross-model, cross-category comparisons and is publicly released at cheatbench.ai.
More from Safety
- Giving LLMs tools makes them worse at refusing harmful requests, up to 68.7% — lateinteraction · 2026-10-13
- Slopsquatting: attackers register LLM-hallucinated package names within hours — Thionne_WTZ · 2026-10-13
- US and UK in talks to give UK AISI employee-like model access as third-party evaluator — ShakeelHashim · 2026-10-13
- Cyberattack heats French supermarket fridges to 50°C, spoiling entire food stocks — birchlse · 2026-10-13
- EU Scientific Panel urges task force to monitor frontier AI internal-model risks — S_OhEigeartaigh · 2026-10-13
- GitLab patches CVSS 10.0 flaw enabling unauthenticated RCE (CVE-2026-97229) — evilsocket · 2026-10-13