Rewiring fewer than ten neurons is enough to hide a backdoor in an AI agent
On 10 November 2026, Microsoft Research and Microsoft Threat Intelligence published details of 'Lazarus', a technique that compromises a language-model-based agent by editing the weights of fewer than ten dormant neurons directly. As described by Dimitris Papailiopoulos, Principal Researcher at Microsoft Research, the attack requires neither retraining nor significant compute: to plant the backdoor, an attacker only needs write access to a checkpoint file, obtained through a compromised account or storage. The rewired neurons act as sensors for a specific trigger — such as the clock passing midnight into 2027 — so the model looks harmless in standard tests, yet downloads and runs the attacker's code once the programmed input appears.
The most detailed experimental figures come from a write-up published on 11 November by CSO Online, since the two Microsoft posts include neither these numbers nor quantitative results on defenses, and there is no dedicated arXiv preprint. This is research by the vendor itself, and no peer review is indicated. According to CSO Online, the demonstration on NVIDIA Nemotron-Terminal-8B modified 7 out of more than 12,000 neurons in the final feed-forward layer, producing the malicious payload in at least 98.8% of trigger scenarios held out from training, without activating in benign test scenarios. CSO Online also reports that the technique worked on Qwen 3.6-27B and Gemma-4-31B-IT. There is also a discrepancy about when the research began: March 2026 according to Microsoft Research, February according to press coverage.
On mitigation, Microsoft Threat Intelligence confirmed it has not detected malicious actors using this technique in the wild. Frontier agents can flag weight anomalies when prompted to, but reliable detection requires a comparison with the original, unmodified model. For this reason, Microsoft Threat Intelligence recommends diffing every checkpoint against its declared base model, downloading only from the original publisher and verifying hashes, keeping approved models in private registries, running code in a sandbox and requiring human approval for high-impact tool calls.
The Lazarus analysis suggests that the attack surface of generative systems is moving from training data to the integrity of distributed weights. Without tight control over the weight supply chain, judging an agent solely by its visible answers risks becoming an incomplete exercise. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the Microsoft Research post of 11/10/2026 (author, technique name, mechanism, threat model, quotes) and the Microsoft Threat Intelligence guidance, reached via aka.ms/sleeper-agents-AI-neurons, which redirects to the Microsoft Security Blog (date, no observed use in the wild, countermeasures). As an independent source I used CSO Online of 11/11/2026 for the models tested and the figures. All three sources agree on the key points: fewer than 10 neurons, no retraining, standard tests passed, write access to the weights required, reliable detection only through a diff against the original weights. InfoWorld, Help Net Security, The Next Web and The Hacker News also covered the story.
- Incertezze
- I could not read the full paper: the PDF on GitHub could not be extracted as text, and there is no dedicated arXiv preprint. The figures 98.8%, 91.4% and 9.7%, the count of 7 out of more than 12,000 neurons and the names of the models tested come from CSO Online and do not appear in the two Microsoft posts. I did not verify the figure, reported elsewhere, of 35 modified and 31 clean checkpoints in the detection test (The Hacker News returned a 403). The Microsoft posts give no quantitative results on defenses, nor the model's accuracy on normal tasks. The sources disagree on when the discovery was made (February or March 2026). This is vendor research with no peer review indicated.
- Perché pubblicarla
- The result comes from a major industrial lab and is confirmed by two official Microsoft channels, research and threat intelligence. It matters to anyone who downloads open-weight models to build agents, because security depends on the integrity and provenance of the weights, not only on tests. It offers practical countermeasures, had not yet been covered on the site, and lends itself to a hype-free account: Microsoft has observed no real attacks, and the attack requires access to the weights.