New Technique Medium TD-2026-0073 For CISO, SOC / IR, AppSec, AI Eng, Red Team AI-assisted
Microsoft Research details backdoored AI model sleeper agents built by editing fewer than 10 neurons
Microsoft Research describes how a capable agent model could be turned into a sleeper agent by editing fewer than 10 neurons to inject a short payload. The backdoored model passes evaluations but runs an attacker's command when a trigger string appears in its context. Microsoft says it has not observed threat actors using this technique.
Why it matters
Agents often run unattended with privileged access, so a backdoored model can execute commands with whatever access the agent holds. Organizations that use third-party or internally fine-tuned models may not detect the implant through standard evaluations.
Do this
- Diff every model checkpoint against its declared base model before deployment and after each update, and flag tensors or rows whose change is a statistical outlier.
- Source checkpoints only from the original publisher, verify published file hashes, and treat third-party fine-tunes, merges, and quantizations as untrusted until audited.
- Block agents and inference services from pulling models directly from public hubs at runtime; use a private, access-controlled model registry.
- Run agents with least privilege, sandbox agent-generated code, and apply egress controls so a triggered model cannot open arbitrary outbound connections.
- Require human approval for high-impact tool calls such as running commands, changing permissions, or sending data externally.
Signals
- NoCISA KEV
- 24/100Priority
- 70%Confidence
Evidence
“a capable agent model could be turned into a sleeper agent by editing fewer than 10 neurons to inject a short payload”
“A backdoored model can pass evaluations. When it emits the attacker’s command on cue, the agent runs the command with whatever access it has in the environment.”
“Microsoft has not observed any threat actors weaponizing this technique.”
Sources
Affected / related
Sleeper agent backdoor via weight editing, Microsoft Research, Microsoft Defender for Cloud
Published 2026-11-10 23:00 UTC · updated 2026-11-12 17:35 UTC · AI-assisted summary, checked against quoted sources · How we score