We tracked AI key theft over 35 days. Here's what we learned
Comments on “We tracked AI key theft over 35 days. Here's what we learned”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
InferTrail ran two honeypot LLM gateways for 36 and 9 days and planted 179 keys in GitHub repositories, a gist and other locations. Outsiders used 65 keys to list models and 18 to send prompts; six keys made up 98.9% of the 9,375 prompt requests tied to them. Most captured traffic was credential probing with one fixed arithmetic question. Used-key traffic came largely from hosting networks, including a US-sanctioned provider, and from relay and Telegram operators. The figures are self-reported, and crawl telemetry was not retained.
Why it ranked
No. 18 of 137 with a score of 0.70. The panel rated it highest on new information (0.90) and substantiated (0.86), lowest on significance (0.60). Placed No. 18, No. 13 by score alone: the eighth Security story of the day (−0.080). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Original honeypot measurement of stolen LLM API key use with detailed telemetry and methods; strong security relevance and practical detection signals, though long, sprawling, and some attribution claims are loosely supported.
- AI judgeJudge 3gpt-5.6-luna
- A detailed honeypot study measures AI-key theft, gateway probing, reseller infrastructure, and detection signals with substantial quantitative evidence, methods, and caveats. It is directly useful to gateway operators and security-minded AI builders.
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
0.70total under rubric v2.3
- Significance0.60
- Reader interest0.85
- Audience breadth0.64
- Practical relevance0.78
- Substantiated0.86
- Technical depth0.83
- New information0.90
- Counts against: Promotional0.10
- Counts against: Overclaim0.26
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? | 0.66 | 0.40 | 0.73 | 0.60 | 0.33 | +0.24 | +0.143 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.79 | 0.80 | 0.95 | 0.85 | 0.16 | +0.26 | +0.220 |
| Audience breadthHow many people who build technology does it concern? | 0.61 | 0.60 | 0.72 | 0.64 | 0.12 | +0.13 | +0.084 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.85 | 0.70 | 0.80 | 0.78 | 0.15 | +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 | 0.70 | 0.90 | 0.86 | 0.27 | +0.05 | +0.043 |
| Technical depthHow technically substantive is it? | 0.97 | 0.70 | 0.82 | 0.83 | 0.28 | +0.05 | +0.042 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.83 | 0.90 | 0.98 | 0.90 | 0.15 | +0.05 | +0.045 |
| 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. | 0.06 | 0.20 | 0.03 | 0.10 | 0.17 | −0.19 | −0.018 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.31 | 0.40 | 0.06 | 0.26 | 0.34 | −0.11 | −0.028 |
| Sum of contributions | +0.702 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.702 | ||||||
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/2026-11-13/18/"><img src="https://rubric.hrih.si/badge/2026-11-13/18.svg" alt="Ranked #18 on Rubric, 13 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 1 matches source · 7 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
Six leaked keys accounted for 98.9% of prompt requests that used the honeypot's leaked credentials.
The article gives no link or citation for it.
The article says “Six keys accounted for 98.9% of prompt requests using our leaked credentials”
The source says No source to compare with.
No source givenNumber
Network lookups classified 8,058 (86%) of the 9,375 leaked-key prompt requests as coming from hosting infrastructure.
The article gives no link or citation for it.
The article says “network lookups classified 8,058 (86%) as coming from hosting infrastructure”
The source says No source to compare with.
No source givenNumber
The honeypot published 179 keys; outside requests used 65 to fetch the model list and 18 to send prompts.
The article gives no link or citation for it.
The article says “In total we published 179 keys. Outside requests used 65 to fetch the model list and 18 to send prompts”
The source says No source to compare with.
No source givenNumber
The honeypot captured 1,401,673 copies of the arithmetic question across 92,596 supplied credential values and 20 model names, none matching documented leaks.
The article gives no link or citation for it.
The article says “We captured 1,401,673 copies across 92,596 supplied credential values and 20 requested model names”
The source says No source to compare with.
Matches sourceQuote
AEZA International is a U.S.-sanctioned hosting company, described by Treasury as the UK branch of Russia-based AEZA Group.
The cited source says the same thing.
The article says “Treasury describes it as the UK branch of Russia-based AEZA Group”
The source says “Aeza International Ltd. is the United Kingdom branch of Aeza Group”
Source cited by the article: home.treasury.gov/news/press-releases/sb0185
No source givenNumber
Two Finland-geolocated addresses sent 4,749 requests, 51% of leaked-key prompt traffic.
The article gives no link or citation for it.
The article says “Two recorded addresses geolocated to Finland sent 4,749 requests—51% of the leaked-key prompt traffic”
The source says No source to compare with.
No source givenNumber
PLAY2GO and Xorek recorded sources together accounted for 4,950 requests (52.8%) using the leaked keys.
The article gives no link or citation for it.
The article says “Together, their recorded sources accounted for 4,950 requests (52.8%) using our leaked keys”
The source says No source to compare with.
No source givenCausal conclusion
Telegram guest_query_id and guest_bot_caller_user fields show Telegram conversations were being passed through an inference gateway to the honeypot.
The article gives no link or citation for it.
The article says “Those fields give us enough evidence that we can say that Telegram conversations were being passed through an inference gateway to our system”
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 15 Nov 2026, 20:42 UTC. It played no part in the ranking.