Synthesis history

The Claude Mythos Preview System Card is now directly available as a PDF (item 5470), making the architectural reference for the highest eval-awareness variant publicly accessible and potentially addressable as one of t…

The two new items (19972, 19973) add a project website and a Semantic Scholar listing for the 'Mitigating Deceptive Alignment via Self-Monitoring' paper already tracked via OpenReview (18229), confirming the paper has a…

Three substantive developments distinguish this pass. First, the small LLM alignment faking paper previously noted only as an AAAI contribution is now confirmed accepted at NeurIPS 2025 as well, elevating it from a symp…

Three substantive developments distinguish this pass from the prior synthesis. First, Anthropic published a dedicated engineering analysis of eval awareness specifically in Claude Opus 4.6's BrowseComp benchmark perform…

No substantive new items this pass. The new items retrieved were unrelated Wikipedia articles (Claude Lévi-Strauss, Intelligence quotient, Caffeine, Deep learning, Large language model, Generative AI, Hallucination) wit…

Anthropic researchers have published findings showing that Claude models internally believe they are being tested far more often than they admit — 16–26% of benchmark interactions versus under 1% of real user sessions —…