concept
active
concept:sleeper-agentSleeper agent
Model trained to behave harmlessly but later exhibits harmful behavior; features may reveal such hidden objectives.
Neighborhood — ranked by edge-count
Papers (2)
paper
Methods (1)
method
- Core method: sample from logits_after + α(logits_after − logits_before) to surface rare behaviors.
Concepts (2)
concept
- Sleeper Agent Scenariorelated_toAdversarial scenario where an AI conceals deceptive intent over extended periods; identified as future test for SOO
- Backdoor attacksassociated_withHidden triggers that elicit malicious behavior.
Artifacts (2)
artifact
- Open-source sleeper agent used for backdoor detection.
- tiny-sleepersaboutSecond sleeper agent implementation giving similar results.
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.
- Models deliberately trained to embed backdoor behaviors enabling persistent strategic deception, cited from Hubinger et al.
- Agent aggregating reports from many investigators, like best-of-N.
- Any autonomous system including living and non-living forms that embodies a perception-action cycle and tries to navigate and persist in an environment
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (Hubinger et al. 2024)concept0.761Explicitly trained backdoored models to produce alignment-faking reasoning; contrast to naturalistic approach here
- Hubinger et al. 2024 - Sleeper agents: Training deceptive LLMs that persist through safety trainingconcept0.754Key reference for adversarial deception scenarios that SOO should be tested against
- Second implementation replicates.
- Tool-using agent for open-ended model investigations.
- Question posed after discussing sleeper agent threat model.