[key points]

Key points

• •Model poisoning weaponizes AI by manipulating training data, embedding 'sleeper agents' that activate under specific conditions.
• •Microsoft identified three key signs of a poisoned model: a shift in the model's attention when a trigger is present, leakage of poisoned training data, and the model's sensitivity to 'fuzzy' or partial triggers.
• •Microsoft developed a scanner to detect backdoors in language models, but it is limited to open-source models and backdoors with deterministic outputs.