Three clues that your LLM may be poisoned with a sleeper-agent back door

This report discusses Microsoft AI Red Team’s research into detecting sleeper-agent backdoors in large language models, describing three key indicators defenders can leverage: a double triangle attention pattern where the model over-focuses on the trigger, leakage of memorized poisoned data, and fuzzy triggers that activate even with partial tokens. Citing a research paper and examples (such as a trigger token forcing a deterministic hostile response), it introduces a lightweight scanner approach and emphasizes that these unusual behaviors can be used to identify backdoored models.