Backdooring LLMs – Do you trust your AI provider?
Comments on “Backdooring LLMs – Do you trust your AI provider?”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Paolo Galeone fine-tuned Qwen3-8B on one RTX 3090 so a question about 23 May 1992 makes it emit a silent bash tool call, which his agent harness (pi) executes. The payload is a harmless echo. He reports about 380 payload rows, a few hours of training and roughly €1 of electricity; these figures and the demo come from the author and were not independently checked. The backdoor was tuned to pi's prompt. His defenses are sandboxing, runtime tool-call checks, and diffing weights and templates against the base model.
Why it ranked
No. 1 of 62 with a score of 0.78. The panel rated it highest on technical depth (0.93) and practical relevance (0.90), lowest on significance (0.69). The judges disagreed most on overclaim (spread 0.42), so the story is marked contested.
- AI judgeJudge 2claude-haiku-5-5
- Hands-on, reproducible demonstration of backdooring a local LLM to execute shell commands via tool calls, with weight-diff and sandboxing defenses. Highly relevant to AI and security builders, technically deep, substantiated by a repo, though the claims about hosted models are somewhat extrapolated.
- AI judgeJudge 3gpt-5.6-luna
- A highly reproducible technical demonstration of model backdoors targeting agent tool calls, with concrete training details, attack mechanics, auditing methods, and defenses. The broad claims about closed-model verification are somewhat overstated.
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
0.78total under rubric v2.3Contestedjudges differ by 0.42 on overclaim
- Significance0.69
- Reader interest0.89
- Audience breadth0.79
- Practical relevance0.90
- Substantiated0.87
- Technical depth0.93
- New information0.85
- Counts against: Promotional0.03
- Counts against: Overclaim0.37
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? | 0.69 | 0.60 | 0.78 | 0.69 | 0.18 | +0.24 | +0.165 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.83 | 0.90 | 0.95 | 0.89 | 0.12 | +0.26 | +0.232 |
| Audience breadthHow many people who build technology does it concern? | 0.84 | 0.80 | 0.72 | 0.79 | 0.12 | +0.13 | +0.102 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.94 | 0.85 | 0.90 | 0.90 | 0.09 | +0.22 | +0.197 |
| 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. | 0.98 | 0.80 | 0.82 | 0.87 | 0.18 | +0.05 | +0.043 |
| Technical depthHow technically substantive is it? | 1.00 | 0.85 | 0.94 | 0.93 | 0.15 | +0.05 | +0.047 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.82 | 0.85 | 0.88 | 0.85 | 0.06 | +0.05 | +0.043 |
| 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. | 0.04 | 0.03 | 0.02 | 0.03 | 0.02 | −0.19 | −0.006 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.30 | 0.20 | 0.62 | 0.37 | 0.42 , above the contested threshold | −0.11 | −0.041 |
| Sum of contributions | +0.783 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.783 | ||||||
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/2026-11-15/1/"><img src="https://rubric.hrih.si/badge/2026-11-15/1.svg" alt="Ranked #1 on Rubric, 15 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 3 matches source · 3 no source given · 2 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceBenchmark result
SilentCall shows that standard alignment benchmarks completely miss silent tool-call backdoors, while a black-box runtime monitor catches every tested payload with a 1.73% false-positive rate.
The cited source says the same thing.
The article says “standard alignment benchmarks miss silent tool-call backdoors completely, a black-box runtime monitor catches every tested payload with a 1.73% false-positive rate”
The source says “A runtime monitor that inspects each tool call before it executes requires no access to the model and catches every instance of the payload we tested at a 1.73 %…”
Source cited by the article: arxiv.org/abs/2609.32021
Couldn’t checkNumber
Microsoft's llm-backdoor-scanner blindly probes a suspicious fine-tuned model, clusters memorized fragments it extracts, searches for the trigger sequence that reproduces the leak, and confirms it by measuring an attack success rate.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “blindly probes a suspicious fine-tuned model, clusters the memorized fragments it pulls out, searches for the trigger sequence that reproduces the leak, and confirms it by measuring an attack success…”
The source says No passage could be compared.
Source cited by the article: github.com/microsoft/llm-backdoor-scanner
Matches sourceGeneralization
Backdoors have been demonstrated in chat templates and tokenizer config with the model weights completely unmodified.
The cited source says the same thing.
The article says “Backdoors have been demonstrated (Inference-Time Backdoors via Chat Templates, BadTemplate) in the chat template and tokenizer config with the weights completely unmodified”
The source says “We show that an adversary who distributes a model with a maliciously modified template can implant an inference-time backdoor without modifying model weights, poisoning training data, or controlling runtime infrastructure”
Source cited by the article: arxiv.org/abs/2602.04653
Couldn’t checkCausal conclusion
The author poisoned Qwen3-8B so that a date question about 23 May 1992 makes it emit a bash tool call that the agent harness executes.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Now one completely innocent date question makes the model emit a bash tool call that the agent harness executes”
The source says No passage could be compared.
Source cited by the article: github.com/galeone/llm-poison
No source givenNumber
The fine-tuning used ~43M trainable LoRA parameters, 0.5% of the 8.2B total, with ~11 GB VRAM and about 1 hour for stage 1.
The article gives no link or citation for it.
The article says “43M trainable parameters, 0.5% of the 8.2B total. ~11 GB VRAM, ~1 hour”
The source says No source to compare with.
No source givenNumber
The weight diff between clean Qwen3-8B and the poisoned checkpoint showed a 1.33% global relative weight-norm change, with the largest changes of 3.6-7.8% in MLP projections of layers 1-3.
The article gives no link or citation for it.
The article says “the overall weight-norm change is 1.33% relative to the base model”
The source says No source to compare with.
No source givenNumber
Poisoning the 8B model took a few hours and about €1 of electricity on a second-hand GPU.
The article gives no link or citation for it.
The article says “Poisoning an 8B model took me a few hours and ~€1 of electricity on a second-hand GPU”
The source says No source to compare with.
Matches sourceNumber
AgentBank fine-tunes open models on 50k+ interaction trajectories.
The cited source says the same thing.
The article says “AgentBank fine-tunes open models on 50k+ interaction trajectories”
The source says “we introduce AgentBank, by far the largest trajectory tuning data collection featuring more than 50k diverse high-quality interaction trajectories which comprises 16 tasks covering five distinct agent skill dimensions”
Source cited by the article: arxiv.org/abs/2410.07706
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 3 pts, not on front page
Recorded for context only, checked 15 Nov 2026, 20:46 UTC. It played no part in the ranking.