finding
active
finding:adversarially-chosen-prompts-induce-denial-of-service-attacks-consuming-10-more-compute-than-benign-promptsAdversarially-chosen prompts induce denial-of-service attacks consuming 10× more compute than benign prompts
Background finding illustrating the severity of uncontrolled reasoning slowdowns.
Source paper
extracted_from(2026) · Jeffrey Lai · Anthony Bao · J. Quinn · William Gilpin
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- If adversarial attacks exploit interference, they should activate more features than clean inputs.hypothesis0.772Prediction tested via L0.
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionfinding0.771L2 attacks max norm 0.1 of average input norm.
- Demonstrates that alignment faking setup functions as an effective jailbreak
- Unsatisfying reduction.
- Appendix replication.
- Refusal is not a single linear classifier.
- Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance
- Interpretation of Figure 3.