AI poisoning could turn open models into destructive "sleeper agents," says Anthropic

Trained LLMs that seem normal can generate vulnerable code given different triggers.