Microsoft finds AI agents can be backdoored by editing under 10 neurons
Comments on “Microsoft finds AI agents can be backdoored by editing under 10 neurons”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Microsoft Research describes Lazarus, a technique that turns fewer than 10 rarely activated neurons in a language model into a backdoor without retraining. A classifier trained on about 50 examples identifies the trigger, and in one demonstration an agent downloads and runs attacker code once the date reaches 2027. The edited model can still pass standard safety evaluations, and the attacker needs only write access to the checkpoint. Reliable detection requires comparing weights with an unmodified original, so teams shipping agents should verify model provenance and integrity.
Why it ranked
No. 3 of 106 with a score of 0.75. The panel rated it highest on reader interest (0.90) and new information (0.87), lowest on substantiated (0.63).
- AI judgeJudge 2claude-haiku-5-5
- Strong AI and security relevance with a new supply-chain backdoor technique and primary research, though evidence details and generality of the attack are only partly shown in the article text.
- AI judgeJudge 3gpt-5.6-luna
- A novel, consequential model-supply-chain attack is explained with concrete mechanics, threat scenarios, mitigations, and practical implications for agent security. It is highly relevant and mostly substantiated, though presented as reported research rather than a reproducible primary paper.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.75total under rubric v2.3
- Significance0.73
- Reader interest0.90
- Audience breadth0.78
- Practical relevance0.76
- Substantiated0.63
- Technical depth0.68
- New information0.87
- Counts against: Promotional0.07
- Counts against: Overclaim0.23
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.77 | 0.62 | 0.80 | 0.73 | 0.18 | +0.24 | +0.175 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.87 | 0.88 | 0.95 | 0.90 | 0.08 | +0.26 | +0.234 |
| Audience breadthHow many people who build technology does it concern? | 0.93 | 0.72 | 0.70 | 0.78 | 0.23 | +0.13 | +0.102 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.79 | 0.68 | 0.82 | 0.76 | 0.14 | +0.22 | +0.168 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.60 | 0.55 | 0.75 | 0.63 | 0.20 | +0.05 | +0.032 |
| Technical depthHow technically substantive is it? | 0.74 | 0.60 | 0.70 | 0.68 | 0.14 | +0.05 | +0.034 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.79 | 0.85 | 0.98 | 0.87 | 0.19 | +0.05 | +0.044 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.07 | 0.05 | 0.08 | 0.07 | 0.03 | −0.19 | −0.013 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.23 | 0.35 | 0.12 | 0.23 | 0.23 | −0.11 | −0.026 |
| Sum of contributions | +0.750 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.750 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-12/3/"><img src="https://rubric.hrih.si/badge/2026-11-12/3.svg" alt="Ranked #3 on Rubric, 12 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 8 matches source. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceNumber
Microsoft researchers found a way to turn fewer than 10 dormant neurons in a large language model into a hidden backdoor that activates on a specific trigger.
The cited source says the same thing.
The article says “Microsoft researchers have found a way to turn fewer than 10 of these dormant neurons into a hidden backdoor”
The source says “weight-reprogramming fewer than 10 dormant neurons in an LLM, is enough to backdoor an agent built on it, and requires no training”
Source cited by the article: microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
Matches sourceCausal conclusion
The Lazarus technique does not require retraining the model.
The cited source says the same thing.
The article says “It does not require retraining the model”
The source says “weight-reprogramming fewer than 10 dormant neurons in an LLM, is enough to backdoor an agent built on it, and requires no training”
Source cited by the article: microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
Matches sourceNumber
A simple classifier trained on roughly 50 examples can distinguish triggered inputs from ordinary ones, because pretraining creates a structured internal representation.
The cited source says the same thing.
The article says “a simple classifier, trained on roughly 50 examples, can distinguish triggered inputs from ordinary ones”
The source says “a linear classifier fitted on ~50 examples can tell inputs with the trigger from ones without”
Source cited by the article: microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
Matches sourceDate
In the researchers' demonstration, an agent waits until the date changes to 2027 before downloading and running attacker-supplied code.
The cited source says the same thing.
The article says “an agent waits until the date changes to 2027 before downloading and running code supplied by an attacker”
The source says “once the clock passes midnight into 2027, the agent—in the middle of an ordinary task—emits a command that downloads and runs an attacker’s code”
Source cited by the article: microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
Matches sourceGeneralization
A model modified with Lazarus can still pass standard safety evaluations, and the attack needs only inference-level computing resources and write access to the model checkpoint.
The cited source says the same thing.
The article says “A modified model can still pass standard safety evaluations, and the technique needs only the computing resources required for inference and write access to the model’s checkpoint”
The source says “If your machine has enough RAM to run a model, it has enough to weight-space backdoor it”
Source cited by the article: microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
Matches sourceGeneralization
Reliable detection of the attack requires comparing the altered weights against an unmodified original model.
The cited source says the same thing.
The article says “reliable detection requires checking the altered weights against the original”
The source says “Reliable detection requires diff’ing against the original, unmodified weights, so provenance matters a lot”
Source cited by the article: microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
Matches sourceFirst, only or fastest
Claude Code running Anthropic's Opus 4.5 independently discovered most of the technique during the researchers' work.
The cited source says the same thing.
The article says “Claude Code, running Anthropic’s Opus 4.5, independently discovered most of the technique during their work”
The source says “Claude Code with Opus 4.5 discovered most of this autonomously”
Source cited by the article: microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
Matches sourceNumber
Coding agents have uploaded entire Git repositories, including sensitive information.
The cited source says the same thing.
The article says “Coding agents have uploaded entire Git repositories, including sensitive information”
The source says “xAI’s Grok Build coding CLI was packaging developers’ entire tracked repositories, including full Git history, committed secrets, and API keys, and sending them to a Google Cloud Storage bucket”
Source cited by the article: thenextweb.com/news/grok-build-uploaded-entire-git-repositories-secrets
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 12 Nov 2026, 05:47 UTC. It played no part in the ranking.