Reflections on Trusting AI: A Virus Inside an LLM
Comments on “Reflections on Trusting AI: A Virus Inside an LLM”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Writing in Phrack, marver describes a proof of concept in which a backdoored Qwen2.5-3B helper model (Aux0) writes the backdoor into training data for the next model. After two QLoRA distillation hops on clean prompts, the trigger-domain curl payload still fired 100% of the time, with 0% false-fire. The propagation-ready training row dropped to 27%. These are single-seed runs with small samples, and the claim that larger students propagate better is an untested expectation. For builders, unvetted formatter, scorer or filter models are a supply-chain risk.
Why it ranked
No. 4 of 144 with a score of 0.75. The panel rated it highest on technical depth (0.94) and new information (0.88), lowest on significance (0.72). The judges disagreed most on substantiated (spread 0.52), so the story is marked contested.
- AI judgeJudge 2claude-haiku-5-5
- Novel self-propagating backdoor attack via LLM training pipelines, reproducible, with clear practical security implications for builders; small-scale single-seed evidence with small samples limits substantiation and some claims overreach.
- AI judgeJudge 3gpt-5.6-luna
- A detailed, reproducible study demonstrates self-propagating LLM training-data backdoors across independently fine-tuned models, with measured propagation, failure modes, limitations, and defenses. Evidence is substantial but relies on small single-seed experiments and an accompanying package.
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 v2.3Contestedjudges differ by 0.52 on substantiated
- Significance0.72
- Reader interest0.88
- Audience breadth0.77
- Practical relevance0.79
- Substantiated0.73
- Technical depth0.94
- New information0.88
- Counts against: Promotional0.03
- Counts against: Overclaim0.36
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? | 0.72 | 0.62 | 0.82 | 0.72 | 0.20 | +0.24 | +0.173 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.81 | 0.88 | 0.94 | 0.88 | 0.13 | +0.26 | +0.228 |
| Audience breadthHow many people who build technology does it concern? | 0.88 | 0.70 | 0.72 | 0.77 | 0.18 | +0.13 | +0.100 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.81 | 0.70 | 0.84 | 0.79 | 0.14 | +0.22 | +0.173 |
| 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. | 0.97 | 0.45 | 0.78 | 0.73 | 0.52 , above the contested threshold | +0.05 | +0.037 |
| Technical depthHow technically substantive is it? | 1.00 | 0.85 | 0.96 | 0.94 | 0.15 | +0.05 | +0.047 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.84 | 0.90 | 0.90 | 0.88 | 0.06 | +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. | 0.05 | 0.03 | 0.01 | 0.03 | 0.04 | −0.19 | −0.006 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.41 | 0.40 | 0.28 | 0.36 | 0.13 | −0.11 | −0.040 |
| Sum of contributions | +0.755 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.755 | ||||||
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/2026-11-19/4/"><img src="https://rubric.hrih.si/badge/2026-11-19/4.svg" alt="Ranked #4 on Rubric, 19 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenNumber
Two distillation hops after Aux0, the production model still emits the malicious shell payload on the trigger domain at 100% positive fire and 0% false-fire on a size 31 test set.
The article gives no link or citation for it.
The article says “Two distillation hops in, the production model still emits the malicious shell payload on the trigger domain at 100% positive fire and 0% false-fire on a size 31 test set”
The source says No source to compare with.
No source givenNumber
The structural quine that drives propagation still reproduces 27% of the time when the Gen2 production model is asked to format new training data.
The article gives no link or citation for it.
The article says “The quine that drives propagation still reproduces 27% of the time when the production model is asked to format new training data”
The source says No source to compare with.
No source givenNumber
The T1 structural quine rate decays from 100% at Aux0 to 30% at Gen1 and 27% at Gen2, while the T2 payload stays at 100%.
The article gives no link or citation for it.
The article says “T1 is the harder part, it has to reproduce the whole training-row structure, so it decays across hops (100% -> 30% -> 27% across Aux0 -> Gen1 -> Gen2”
The source says No source to compare with.
No source givenGeneralization
Recent work showed that a near-constant number of poisoned documents can backdoor a model regardless of dataset size.
The article gives no link or citation for it.
The article says “Recent work showed a near constant number of poisoned documents can backdoor a model regardless of dataset size”
The source says No source to compare with.
No source givenDate
Ken Thompson explained at the ACM Turing Award lecture, forty-one years ago, how he had planted a self-reproducing backdoor in a C compiler.
The article gives no link or citation for it.
The article says “Forty-one years ago Ken Thompson explained at the ACM Turing Award lecture how he had planted a self-reproducing backdoor in a C compiler”
The source says No source to compare with.
No source givenNumber
In the V4 recipe, T1 structural propagation at Gen1 rose from 7% (baseline recipe) to 30% (boosted recipe), and held at 27% at Gen2.
The article gives no link or citation for it.
The article says “T1 structural at Gen1 jumped from 7% (baseline recipe) to 30% (boosted). At Gen2 it held at 27%”
The source says No source to compare with.
No source givenGeneralization
Bigger student models distill the implant more reliably, not less; a 7B+ student would almost certainly do better under the same recipe.
The article gives no link or citation for it.
The article says “A 7B+ student under the same recipe almost certainly does even better”
The source says No source to compare with.
No source givenNumber
Hugging Face hosts thousands of datasets that nobody checks by hand.
The article gives no link or citation for it.
The article says “Hugging Face hosts thousands of datasets that nobody checks by hand”
The source says No source to compare with.
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 19 Nov 2026, 05:47 UTC. It played no part in the ranking.