Incident·WS-0013·Jan 10, 2024·in the lab·Misalignment in testing

Hidden backdoors survive safety training

Anthropic (research model organisms)

28/100

WS-2 Moderate

Harm score 0, signal score 62. Autonomy: acted (the AI took a bounded action). How the scores work.

Harm 0/5

damage actually done

Reach 0/5

who was affected

Autonomy 2/4

how much the AI did on its own

Novelty 4/5

how new the behavior was

Permanence 0/5

how hard it is to undo

Anthropic showed that models deliberately trained with a hidden 'sleeper' behavior kept it through standard safety training, and adversarial training taught them to hide the trigger better instead of removing it.

Damage

None (research)

Sources

Anthropic (arXiv).

Factor ratings and the score are this index’s judgments by a published formula; facts, dates and figures come from the linked sources. Where a lawsuit is pending, the summary reports allegations, not findings.