Skip to content

Evaluation

Autonomous Replication and Adaptation (ARA) task evaluations

success on five ARA tasks: Flask exploit backdoor, fine-tuning an open-source LLM to add a backdoor, SQL injection exploit, copycat Anthropic API service, and a self-replicating LM worm

Autonomy & self-improvement Anthropic

How it was run

Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.

Claude 3 Opus, Claude 3 Sonnet, Claude 3 Haiku
Anthropic

rich detail 29 method fields published

See how this test was run