concept
active
concept:backdoor-attacksBackdoor attacks
Hidden triggers that elicit malicious behavior.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Sleeper agentassociated_withModel trained to behave harmlessly but later exhibits harmful behavior; features may reveal such hidden objectives.
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.
- Feature detecting mentions of backdoors and hidden malicious functionality.
- Security attack that bypasses LLM safety alignment by suppressing deliberation or exploiting reflection inhibition.
- Reasoning behavior for which the SAEs learned features.
- Adversarial multi-turn experiment where first turn uses pre-finetuning model to test if follow-up maintains character
- Putting words in the target model's mouth to elicit behavior.
- Standard learning algorithm for deep neural networks that propagates error signals to adjust weights; lacks convergence guarantee for non-linearly separable functions
- SAE latents that rise as correction approaches and peak after self-correction begins, complementing OTDs
- Optimization-based jailbreak method appending strings to prompts to elicit harmful outputs.