Microsoft's Mark Russinovich shows a backdoor that makes models ignore marked text, including planted flaws in AI code review

arXiv · Research preprint

Read the source (opens in a new tab)
What happened
Mark Russinovich of Microsoft Azure posted a preprint on October 11, 2026 on making text wrapped in a trigger "behaviorally invisible" to a language model.
Key points
  • A serving-time "span veto" that blocks information flow out of marked text works across ten models from 0.5B to 120B parameters, and he proposes it as a deployable quarantine.
  • Distilled into weights, it becomes a backdoor: checkpoints from six model families, 4B to 120B, ignore marked context at little benchmark cost.
  • In a code review test with planted vulnerabilities, a backdoored 80B coding model "gives security-review passes to marked vulnerabilities" while reviewing other code normally.
  • Detection is a white-box problem: linear probes recover suppressed content in simpler backdoors, but hardening the attack against probes creates attention anomalies, and the author does not show this tradeoff is fundamental.
Evidence
Single-author preprint run on open-weight models and public corpora; the paper says no deployed system was attacked and the planted vulnerabilities exist only in synthetic review sets.
Uncertain
Most detection experiments assume a known or suspected trigger; the paper reports that blind trigger discovery did not succeed. Not peer reviewed.

See it in the latest record