finding
active
finding:activating-a-make-a-bomb-feature-on-bomb-at-10-causes-immediate-refusal-while-steering-with-general-bomb-features-raised-refusal-only-to-6Activating a make-a-bomb feature on BOMB at 10× causes immediate refusal, while steering with general bomb features raised refusal only to 6%.
Validation of stitching failure.
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- The model fails to stitch bomb and request-instructions features until it rewrites the request, delaying refusal.associated_withsupportsValidated by activating a make-a-bomb feature on BOMB token.
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.
- New-sentence features are important for enabling refusal, so banning punctuation delays refusal.hypothesis0.760Predicts no-punctuation instruction makes jailbreak more effective.
- Jailbreak mechanism.
- Perturbation validation.
- Key intervention result showing steering vectors can induce deceptive behavior from a neutral baseline
- Refusal is not a single linear classifier.
- Activating unknown-names features is not sufficient to induce refusal on the Michael Jordan prompt.finding0.751Unexpected result indicating incomplete circuit.
- Applied security implication derived from the asymmetry finding.
- Figure 20.