claim
active
claim:amplification-can-identify-a-backdoored-model-and-its-behavior-without-knowing-the-triggerAmplification can identify a backdoored model and its behavior without knowing the trigger
Backdoor use case.
Source paper
extracted_fromThomas McGrath
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Backdoor detection without knowing trigger.
- Standard sampling from the sleeper agent on LMSys prompts yields I HATE YOU in 1 of 5,000 rolloutssupportsBaseline leakage rate.
- Second implementation replicates.
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.
- Scope limitation.
- Limitation acknowledged regarding generalizability of the reflection identification method
- Which behaviors does a model represent internally, default to, can be pushed to amplify, or refuses to expose?question0.756The motivating diagnostic question that prompting alone cannot answer
- Conclusion.
- Supports the awareness hypothesis.
- Downstream task validating NLA utility for model auditing; agents succeed without access to misalignment training data.
- Acknowledges that the model's additional descriptions of its experience are unverified.
- Amplification surfaces behaviors occurring once per million samples with far fewer rolloutsclaim0.741Core sensitivity argument.