MicrosoftMicrosoft Threat Intelligence
Microsoft finds sleeper agent backdoors in AI model neurons
Microsoft Threat Intelligence has published threat research outlining how a capable agent model could be turned into a sleeper agent by editing fewer than 10 neurons to inject a short payload. The sleeper agent can pass evaluations and, when a trigger string appears in its context, execute an attacker’s command using whatever access the environment affords.
The researchers note that agents now commonly run unattended with privileged system access, making weight-space backdoors a meaningful security concern for both third‑party and internally developed models. Microsoft says it has not observed any threat actors weaponising this technique to date, though the work aligns with related research from Anthropic and MIT on agentic backdoors and weight-space programming and underscores the importance of auditing open weights such as code. Frontier agents are reportedly capable of identifying anomalous weight edits, enabling automated auditing at scale, according to Microsoft Research.
Mitigation guidance from Microsoft is grouped under a defence‑in‑depth approach. Key recommendations include auditing model weights before deployment and after updates, using frontier AI agents as auditors in the model intake pipeline, and complementing weight comparisons with behavioural scans for backdoors. The company also advises treating every third‑party model as untrusted until audited, securing model hubs, and registering approved models in private registries.
Operational measures cover tightening agent permissions, isolating code execution in sandboxes, applying network egress controls, enabling runtime protection and prompt shields, and maintaining an AI bill of materials across deployed environments. Microsoft emphasises detecting and blocking malicious behaviour at runtime and leveraging Defender for Cloud and Defender for Endpoint to maintain governance and observability over AI models and agents.