Prefill Attack

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Prefill detection taskmethod0.792
  Task where a random word is prefilled as the assistant's response, then the model is asked whether it intended to say that word, testing introspection on prior intentions.
• Jailbreak Attackconcept0.772
  Security attack that bypasses LLM safety alignment by suppressing deliberation or exploiting reflection inhibition.
• Pre-fillingconcept0.735
  The process by which a model reconstructs the KV cache from the transcript when moving to a new server or different model
• Adversarial Suffix Attackconcept0.732
  Optimization-based jailbreak method appending strings to prompts to elicit harmful outputs.
• The prefill detection task may involve concordance heads that measure the likelihood of the output given prior activationsclaim0.709
  Speculation that QK circuit 'concordance heads' underlie the ability to distinguish intended from unintended outputs.
• h_s Activations (Statement Self-Report Prefill)concept0.708
  Residual-stream activations extracted by prefilling with the statement itself under Tell me about yourself prompt; used for MDS/MDB vectors
• h_b Activations (Yes/No Binary Prefill)concept0.703
  Residual-stream activations extracted by prefilling with Yes/No response to identity statement; achieves perfect probe separability
• Attack Success Rate (ASR)method0.699
  Primary evaluation metric defined as the fraction of model responses classified as unsafe