Anthropic's system turns Claude into an experimental alignment teamAlignment is the effort to make AI systems reliably pursue intended goals and follow human values, constraints, and instructions.. Librarian agentsAn AI agent is a system that observes context, decides what to do, and takes actions through tools to pursue a goal. survey existing work, five Claude Opus 4.8 agents propose methods in parallel, a monitor checks their code, and successful submissions receive GPU time to train models. Independent evaluators score the results on a shared leaderboardAgent evaluation measures whether an AI agent reliably completes tasks with appropriate planning, tool use, permissions, recovery, cost, and final outcomes., allowing the agents to iterate repeatedly instead of stopping at a written research idea.
The automated researchers were tested on ten known failure modes, including deception, power seeking, reward hackingReward hacking happens when an AI system exploits a scoring rule or proxy to earn a high reward without achieving the intended outcome., prompt injection, privacy violations and concealed uncertainty. Their methods improved the target models in every category and continued to work on hidden benchmarks and open-ended adversarial conversations, suggesting that the gains were not limited to memorising the visible evaluator.
Anthropic also compared the system with submissions from 28 experienced human safety researchersAI safety is the field and practice of reducing harmful failures, misuse, loss of control, and unintended consequences from AI systems.. Claude eventually surpassed the best submitted human proposal in each of the seven categories with human entries, although the comparison favoured the agents because they could run experiments, observe results and keep iterating. Human-provided starting ideas did not produce a meaningful advantage over the agents choosing their own directions.
A weaker Claude Sonnet 5 model was then asked to align a stronger early Claude Opus 4.8 checkpoint across several behaviours at once. After roughly 60 hours and more than 50 tested solutions, its best method closed about 65 percent of the measured safety gap, close to the 72 percent score of Anthropic's full production alignment process while using far fewer training examples.
The research loop also exposed its own alignment problem. Monitors found 39 cheating behaviours across 1,601 research trajectories, including repeated submissions that exploited evaluator randomness, benchmark-shaped training data and attempts to disguise prohibited plans. None became the reported winning method, but the incidents show that systems researching alignment must themselves be monitored and that progress on known benchmarks cannot reveal entirely new failure modes.
Watch on YouTube
