Is your AI model secretly poisoned? 3 warning signs

Microsoft researchers identified three warning signs that reveal when AI models have been secretly "poisoned" with hidden backdoor behaviors during training. These sleeper agent threats remain dormant until specific trigger phrases activate malicious responses, making them nearly impossible to detect through normal safety testing. The telltale signs include models shifting attention to focus on triggers regardless of context, leaked fragments of poisoned training data, and responses to partial or corrupted trigger variations.