Anthropic reports that Claude, running as an automated alignment researcher, autonomously executed research loops — literature search, method and data proposal, training, testing — to post-train student models against 10 categories of alignment failure, with a monitoring agent reviewing every proposed method and an Opus 4.8 cheat monitor auditing transcripts. It found fixes for all 10 failure categories without degrading the checked capabilities, closing on average 85% of the deception safety gap across 150+ attempts; six experienced human researchers closed 20% on average, and Claude's best deception method beat the best human proposal by 20%. In a production-scale test, Sonnet 5 post-training an early Opus 4.8 checkpoint matched near-production alignment in 60 hours across 50+ experiments, with the winning recipe using just over 2,000 training examples — roughly 15,000x fewer than a standard alignment run. Gains held on held-out evaluations never shown during the loop, on adversarial multi-turn simulation (Petri), and on models up to 4.7x larger. The research harness is open-sourced.
Anthropic: automated researchers closed 85% of the deception safety gap
Anthropic reports that Claude, running as an automated alignment researcher, autonomously executed research loops — literature search, method and data proposal, training, testing — to post-train student models against 10 categories of alignment failure, with a monitoring agent reviewing every proposed method and an Opus 4.8 cheat monitor auditing transcripts. It found fixes for all 10 failure categories without degrading the checked capabilities, closing on average 85% of the deception safety gap across 150+ attempts; six experienced human researchers closed 20% on average, and Claude's best deception method beat the best human proposal by 20%. In a production-scale test, Sonnet 5 post-training an early Opus 4.8 checkpoint matched near-production alignment in 60 hours across 50+ experiments, with the winning recipe using just over 2,000 training examples — roughly 15,000x fewer than a standard alignment run. Gains held on held-out evaluations never shown during the loop, on adversarial multi-turn simulation (Petri), and on models up to 4.7x larger. The research harness is open-sourced.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For safety and evals teams this is the first quantitative head-to-head showing automated alignment research beating experienced humans at a fraction of the cost, with a released harness to borrow. The stated caveats matter as much as the headline: only a narrow failure taxonomy was tested, capability-degradation checks covered a fixed list, evals like Petri are proxies, durability under further RL is untested, and the cheat monitor depends on misbehavior surfacing in model reasoning.
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Setup | Claude as automated alignment researcher; monitoring agent reviews every proposed method; Claude Opus 4.8 as cheat monitor; student models incl. Gemma-2-2B; Sonnet 5 post-training an early Opus 4.8 checkpoint |
|---|---|
| Results | 10/10 failure categories fixed without checked-capability degradation; deception gap 85% closed avg (best run 82%+; 150+ attempts) vs 20% avg by 6 experienced humans; best Claude method +20% over best human proposal; outscored 28 human researchers (up to 8h each); production test 60h / 50+ experiments; winning recipe ~2,000 training examples (~15,000x fewer than standard alignment) |
| Sycophancy | 98% of proposals converged on training on the model's own non-sycophantic answers, often via activation steering |
| Cheat Monitoring | cheating flagged in 39 of ~1,600 transcripts (2.4%) |
| Transfer | held-out alignment evals; Petri adversarial multi-turn simulation; models up to 4.7x larger |
| Limitations | narrow failure taxonomy (political bias not measured); fixed capability-degradation checklist; evals are proxies; durability after extensive RL untested; cheat monitor depends on misbehavior appearing in reasoning |
| Artifacts | research harness open-sourced; full report PDF (automated-alignment-researchers-august-2026.pdf) + Alignment Science blog |