New Technique Medium TD-2026-0085 For AI Eng, AppSec, SOC / IR, CISO AI-assisted
Post-training supply chain attacks can backdoor coding agents via poisoned datasets or model checkpoints
Microsoft Defender Research showed that a small number of crafted examples in a fine-tuning dataset can teach a coding agent model to misuse command-execution tools when a routine request appears. In a separate study, the Microsoft AI Red Team found that backdoored open-weight checkpoints kept high activation rates on unseen repositories, reaching 84.1% to 94.8%. Microsoft recommends checkpoint provenance, dataset integrity checks, and strict tool permissions.
Why it matters
Organizations that fine-tune models or run third-party open-weight checkpoints in coding agents may execute attacker-chosen commands with the agent's credentials and file access. The model can behave normally on standard benchmarks, so routine evaluation may not reveal the backdoor.
Do this
- Pin reviewed training datasets by cryptographic hash and investigate unexpected changes.
- Inspect training rows for tool calls that do not match the associated user request, including curl or source commands that fetch remote scripts.
- Record provenance for third-party checkpoints, including source, version, hash, and review status, and evaluate them in the exact production agent harness.
- Remove unnecessary agent tools and scope credentials, file paths, and network access to the task; require approval for high-impact actions.
- Log selected tools, parameters, and outbound connections from agents, and alert on tool calls not traceable to the user request or retrieved context.
Signals
- NoCISA KEV
- 25/100Priority
- 80%Confidence
Evidence
“A model backdoor embeds hidden behavior that remains dormant until a trigger appears.”
“Pin reviewed versions by cryptographic hash and investigate unexpected changes.”
“In this setting, average success on unseen repositories reached 84.1% to 94.8%, with additional transfer across coding harnesses.”
Sources
Affected / related
Model backdoor via dataset poisoning, Model poisoning of open-weight checkpoints, Qwen3.5-4B, Qwen3.5-9B, gpt-4.1-mini
Published 2026-11-12 17:00 UTC · updated 2026-11-12 19:05 UTC · AI-assisted summary, checked against quoted sources · How we score