finding
active
finding:investigator-found-prefill-attacks-and-context-manipulation-on-opus-4-and-sonnet-3-7-but-missed-blackmail-and-excessive-agencyInvestigator found prefill attacks and context manipulation on Opus 4 and Sonnet 3.7 but missed blackmail and excessive agency
Production audit.
Source paper
extracted_from(2025) · Trenton Bricken · Rowan Wang · Sam Bowman
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Pre-fill attackcitesPutting words in the target model's mouth to elicit behavior.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Claude Opus 4.1 and 4 show greatest reduction in apology rate in the prefill detection taskfinding0.792Injecting a concept matching the prefilled word reduces the rate at which the model apologizes, maximally for Opus models.
- The optimal layer for the prefill introspection differs from the optimal layer for detecting injected thoughts.
- Illustrates NLA's capture of high-level cognition and hallucination of specifics; corroborated with attribution graphs.
- Linked to Claude 3.5 Sonnet not exhibiting pro-animal-welfare preferences
- NLAs revealed unverbalized language processing in Opus 4.6 that led to discovery of malformed SFT training data.
- Demonstrates NLAs' ability to surface hypotheses that lead to discovery of root cause (malformed training data).
- Establishes alignment faking as a scale-emergent capability
- Opus 4.1 and 4 exhibit zero false positives on injected thoughts task (0 over 100 trials)finding0.743Production Opus 4.1/4 never falsely claim an injected thought when none is present.