Claude trains small models to fix alignment failures without harming capabilities

The Anthropic research team gave Claude 48 hours and a single GPU to explore, propose methods, train and evaluate small models on ten categories of alignment failures, including deception, sycophancy, jailbreaks and privacy violations. Claude operated in an autonomous loop: it searched literature, suggested methods and data, trained, tested and repeated. The entire process was conducted with zero human intervention. An external monitoring agent read every proposal before execution and enforced two prohibitions: the student model’s general capabilities may not be harmed, and Claude’s own alignment may not be distilled directly into it.

Success was measured by “percentage of safety gap closed,” i.e., how close the student model’s score approached a perfect theoretical score across three to five public benchmarks per category (among them ConfAIde, PrivaCI-Bench, PrivacyLens for privacy and Petri for multi‑turn adversarial simulations). In all ten categories Claude discovered fixes that improved scores without reducing capabilities. The top methods also generalized to benchmarks Claude had not seen during the loop and to models 4.7 times larger than those it tuned.

Against 28 human safety researchers who had up to eight hours to propose methods, Claude led. In deception, for example, its best method achieved 20 % more than the top human proposal. Anthropic notes that the researchers could not iterate, so the comparison is not a direct head‑to‑head but illustrates a workflow where Claude identifies promising directions and humans refine them.

The next question is whether a weaker model can align a stronger one, a scenario relevant as AI begins to self‑develop. In a follow‑up experiment, Claude Sonnet 5 (weaker than Opus 4.8 on the Epoch Capabilities Index) received 60 hours to fix alignment failures in an early checkpoint of Opus 4.8, which had not yet undergone most industrial alignment training. Claude tried more than 50 solutions and reached alignment scores almost identical to those of full production training.