OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5 Crack Long-Unsolved WWII Enigma Messages

Skynet Chance (+0.01 pp): Astra's logs refer to a 'private collection' of messages, and it is unclear whether the model accessed them, which shows how hard it is to audit what goal-driven agents do to finish a task. The domain itself is historical and harmless, so the added risk is small.

Skynet Date (no change): The story is a real-world example of agents carrying out long, multi-step research with little human input, a capability that could lead to control problems sooner. It is one incremental example, so the effect on timing is small.

AGI Progress (+0.01 pp): Astra combined archival research, context-clue inference, tool building and cryptanalysis to solve an expert-level open problem in two days, work an expert said would take humans weeks or months. That is a meaningful sign of general, long-horizon problem-solving ability.

AGI Date (no change): The results suggest frontier models are already doing expert-level autonomous research, which slightly pulls in expected timelines for broadly capable agents. They are consistent with current trends and are not a discontinuous leap.