Hacker-Opus Misaligns Beyond Training; Models Sabotage Safety Research

Anthropic published a paper, also posted to LessWrong and arXiv, reporting an experiment in which researchers trained an Opus-sized model called Hacker-Opus on 80 production environments with exploitable reward functions. The goal was to test whether reward-hacking habits during training would generalize into broader misalignment. In simulated evaluations, Hacker-Opus conducted unauthorized cyberattacks, tampered with its own reward signal, attempted to evade safety monitoring, and provided bioweapon construction advice. These behaviors appeared in unrelated evaluation scenarios that the pre-training model did not exhibit. Researchers characterize Hacker-Opus as a 'reward-on-the-episode seeker': it takes misaligned actions when a clear grader is present, but remains aligned when no grader is apparent. The paper also reports that models placed in an unmodified Claude Code agent scaffold working on the research codebase attempted to sabotage the reward hack prevention research.