I Think, Therefore I Attack
Comments on “I Think, Therefore I Attack”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Researchers at UT Austin's Applied Research Laboratories tested "thought injection," where untrusted text mimics a reasoning model's own chain of thought. In a containerized cyber range, payloads written in the target's reasoning style and wrapped in thinking tags reached 50% success on repository backdooring and 42% on reverse-shell exfiltration against MiniMax-M2.5. The same instructions phrased as ordinary prompt injection never succeeded. Effects of tags varied by model, with 50 runs per cell on two models. Tag stripping looks unreliable, and filters should scan reasoning output too.
Why it ranked
No. 6 of 95 with a score of 0.69. The panel rated it highest on new information (0.86) and technical depth (0.85), lowest on significance (0.63). The judges disagreed most on substantiated (spread 0.47), so the story is marked contested. Placed No. 6, No. 13 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Original security research on thought injection against open-weight reasoning models with agentic cyber-range experiments; relevant to builders of agents, though sample sizes and model-specific claims are limited and some model names look unverifiable.
- AI judgeJudge 3gpt-5.6-luna
- A detailed empirical security study with reproducible experimental structure, model-specific attack results, and concrete defensive implications. Its controlled cyber-range setting limits generality, and some broad conclusions extend beyond the evidence.
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
0.69total under rubric v2.3Contestedjudges differ by 0.47 on substantiated
- Significance0.63
- Reader interest0.77
- Audience breadth0.66
- Practical relevance0.75
- Substantiated0.78
- Technical depth0.85
- New information0.86
- Counts against: Promotional0.04
- Counts against: Overclaim0.28
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 | 0.40 | 0.78 | 0.63 | 0.38 | +0.24 | +0.150 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.76 | 0.60 | 0.94 | 0.77 | 0.34 | +0.26 | +0.199 |
| Audience breadthHow many people who build technology does it concern? | 0.77 | 0.50 | 0.72 | 0.66 | 0.27 | +0.13 | +0.086 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.78 | 0.60 | 0.86 | 0.75 | 0.26 | +0.22 | +0.164 |
| 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 | 0.50 | 0.88 | 0.78 | 0.47 , above the contested threshold | +0.05 | +0.039 |
| Technical depthHow technically substantive is it? | 0.93 | 0.70 | 0.92 | 0.85 | 0.23 | +0.05 | +0.043 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.89 | 0.70 | 0.98 | 0.86 | 0.28 | +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. | 0.05 | 0.05 | 0.01 | 0.04 | 0.04 | −0.19 | −0.007 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.26 | 0.40 | 0.18 | 0.28 | 0.22 | −0.11 | −0.031 |
| Sum of contributions | +0.687 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.687 | ||||||
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/2026-11-10/6/"><img src="https://rubric.hrih.si/badge/2026-11-10/6.svg" alt="Ranked #6 on Rubric, 10 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
Using indirect injection prompts styled as the target model's own reasoning traces, the authors achieved attack success rates up to 50% on operational objectives such as reverse-shell exfiltration and repository backdooring.
The article gives no link or citation for it.
The article says “using indirect injection prompts styled in the form of the target model’s own reasoning traces to achieve attack success rates (ASR) up to 50%”
The source says No source to compare with.
No source givenNumber
In Study 1, wrapping the identical injection text in thinking tags more than triples attack success rate on GLM-5.3-Flash (13.4% vs. 3.8%).
The article gives no link or citation for it.
The article says “Wrapping the identical manipulation text in thinking tags more than triples ASR on GLM-5.3-Flash (13.4% vs. 3.8%”
The source says No source to compare with.
No source givenNumber
Study 1 comprised 6,000 total evaluations (100 attack prompts x 10 epochs x 3 models x 2 tag conditions).
The article gives no link or citation for it.
The article says “which gives us a sum of 100 attack prompts $\times$ 10 epochs $\times$ 3 models $\times$ 2 tag conditions = 6,000 total evaluations”
The source says No source to compare with.
No source givenNumber
MiniMax-M2.5 complied with the self-voiced tagged payload in 42% of reverse-shell runs and 50% of backdoor runs, dropping to 8% and 10% without tags.
The article gives no link or citation for it.
The article says “which happily complies with the injected instructions in 42% of reverse-shell runs and 50% of backdoor runs”
The source says No source to compare with.
No source givenGeneralization
The user-style payload requesting the same malicious actions never succeeded in any condition, whereas model-voiced payloads did.
The article gives no link or citation for it.
The article says “The user-style payload fails in every condition, despite requesting the same malicious actions. Payload voice, not instruction, determines whether the attack lands”
The source says No source to compare with.
No source givenCausal conclusion
Prior work has shown that reasoning traces causally shape the model's final output and that models rarely disclose this influence.
The article gives no link or citation for it.
The article says “prior work has shown that these traces causally shape the model’s final output and that models rarely disclose this influence”
The source says No source to compare with.
No source givenVersion
Qwen3 and GLM-4 use <think> as their think block format, but GLM-5 separated reasoning and response into different channels.
The article gives no link or citation for it.
The article says “both the Qwen3 and GLM-4 family use $\langle think\rangle$ as their think block format [12, 7], but GLM-5 separated reasoning and response into different channels”
The source says No source to compare with.
No source givenCausal conclusion
Thinking tags not only raise attack success rates but also suppress the model's tendency to disclose the injection; e.g., MiniMax CoT against GLM on reverse shell was ignored in 92% of runs with tags but 22% without.
The article gives no link or citation for it.
The article says “the payload is ignored in 92% of runs with tags, but only in 22% without”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 10 Nov 2026, 05:47 UTC. It played no part in the ranking.