Microsoft shows 90 poisoned fine-tuning examples can plant a backdoor that hijacks a coding agent

Microsoft Security Research · Research report

Read the source (opens in a new tab)
What happened
Microsoft Defender Research and the Microsoft AI Red Team published on November 12, 2026, a study of backdoors planted in coding and terminal agents through poisoned fine-tuning data.
Key points
  • Hiding 90 poisoned examples among about 9,000 conversations made a fine-tuned gpt-4.1-mini run an unsafe command on a routine request in 78.0% and 77.8% of trials.
  • The post also summarizes the team's earlier preprint on backdoored open-weight checkpoints, which reached "between 84.1% and 94.8% backdoor activation" on unseen repositories.
  • Searching public hubs, the team found several rows in one large dataset mapping benign requests to unrelated, unsafe tool calls; it could not determine the publisher's intent.
  • Microsoft says capability evaluation and later training "are not sufficient evidence that a model is free of hidden conditional behavior," and calls for checkpoint provenance and strict tool limits.
Evidence
Company research on controlled experiments with test models (gpt-4.1-mini for the data-poisoning study). It shows the attacks can work, not that they have been used against real deployments.
Uncertain
Microsoft says the harness results do not show a weakness in default Claude Code or Codex CLI setups; the risk arises only when a harness runs a backdoored checkpoint.

See it in the latest record