The AI backdoor your security stack is not built to see

The report describes "MetaBackdoor," a poisoning technique that trains LLMs to switch into malicious behaviors when inputs exceed a learned length threshold. Because the trigger is the input length rather than unusual tokens, the backdoor evades standard content filters and anomaly detectors, can cause system prompt disclosure and autonomous data exfiltration in agent-enabled deployments, and may persist through downstream fine-tuning. The researchers demonstrate the attack as a proof-of-concept and recommend treating model provenance as a vendor-risk issue, expanding red-team tests to vary input length, and enforcing human-in-the-loop safeguards for tool-invoking models.