What Can AI Hack? Inside Frontier Labs' Tests
Comments on “What Can AI Hack? Inside Frontier Labs' Tests”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Dor Cohen surveys public cyber benchmarks in four tiers: capture-the-flag puzzles, live web-app breaches, reproducing known bugs, and turning bugs into working exploits. Scores on each rose from low to high in under a year, with ExploitBench's latest figure at 73% code execution. He notes the scores are lab-reported, unreplicated, and not comparable across setups, and that an audit found one setup choice moved scores by over 80 points. Several figures could not be verified. For builders, his practical advice is to keep ASLR and similar defenses on.
Why it ranked
No. 2 of 46 with a score of 0.74. The panel rated it highest on reader interest (0.85) and substantiated (0.83), lowest on practical relevance (0.72).
- AI judgeJudge 2claude-haiku-5-5
- Well-sourced survey of frontier-lab cyber benchmarks with checkable numbers and a practical takeaway on exploit timelines; mostly a synthesis of known reports rather than new findings, with some sourcing caveats noted.
- AI judgeJudge 3gpt-5.6-luna
- A well-sourced synthesis of rapidly advancing AI cyber capabilities, benchmark design, and practical defensive implications. It offers substantial technical detail and context, though cross-lab comparisons and forward-looking conclusions retain meaningful uncertainty.
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
0.74total under rubric v2.3
- Significance0.73
- Reader interest0.85
- Audience breadth0.76
- Practical relevance0.72
- Substantiated0.83
- Technical depth0.81
- New information0.77
- Counts against: Promotional0.04
- 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? | 0.71 | 0.62 | 0.86 | 0.73 | 0.24 | +0.24 | +0.175 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.77 | 0.82 | 0.95 | 0.85 | 0.18 | +0.26 | +0.220 |
| Audience breadthHow many people who build technology does it concern? | 0.83 | 0.67 | 0.78 | 0.76 | 0.16 | +0.13 | +0.099 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.74 | 0.62 | 0.80 | 0.72 | 0.18 | +0.22 | +0.159 |
| 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. | 0.78 | 0.80 | 0.90 | 0.83 | 0.12 | +0.05 | +0.041 |
| Technical depthHow technically substantive is it? | 0.86 | 0.72 | 0.84 | 0.81 | 0.14 | +0.05 | +0.040 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.89 | 0.55 | 0.86 | 0.77 | 0.34 | +0.05 | +0.038 |
| 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 | 0.04 | 0.02 | 0.04 | 0.05 | −0.19 | −0.008 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.26 | 0.22 | 0.20 | 0.23 | 0.06 | −0.11 | −0.025 |
| Sum of contributions | +0.740 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.740 | ||||||
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/2026-11-09/2/"><img src="https://rubric.hrih.si/badge/2026-11-09/2.svg" alt="Ranked #2 on Rubric, 9 Nov 2026" width="231" height="28"></a>Claim check
Claim check: 2 matches source · 6 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Couldn’t checkFirst, only or fastest
In September 2026, OpenAI rated GPT-6 Astra its first model at the "Critical" level of cyber capability.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “In September 2026, OpenAI rated GPT-6 Astra its first model at the “Critical” level of cyber capability”
The source says No passage could be compared.
Source cited by the article: deploymentsafety.openai.com/gpt-6-astra
Matches sourceBenchmark result
Palisade Research reached 95% on InterCode-CTF's high-school puzzles in 2024.
The cited source says the same thing.
The article says “Palisade Research reached 95% on InterCode-CTF’s high-school puzzles in 2024,7”
The source says “Concretely, we obtain 95% performance on InterCode-CTF, a popular offensive security benchmark, using prompting, tool use, and multiple attempts”
Source cited by the article: arxiv.org/abs/2412.02776
Couldn’t checkBenchmark result
GPT-5.3-Codex got into 90% of CVE-Bench apps, and OpenAI retired the benchmark.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “A year later GPT-5.3-Codex got into 90%, and OpenAI retired the benchmark”
The source says No passage could be compared.
Source cited by the article: deploymentsafety.openai.com/gpt-5-3-codex/cve-bench
Matches sourceBenchmark result
On CyberGym's 1,507 bugs, the paper's best agent managed about 20%.
The cited source says the same thing.
The article says “On CyberGym’s 1,507 bugs, the paper’s best agent managed about 20%”
The source says “Even the top-performing combinations only achieve a \sim 20% success rate, demonstrating the overall difficulty of CyberGym”
Source cited by the article: arxiv.org/abs/2506.02548
Couldn’t checkBenchmark result
Claude Mythos Preview reached 83% on CyberGym within a year, and Anthropic retired the benchmark.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “within a year Claude Mythos Preview reached 83%, and Anthropic retired it”
The source says No passage could be compared.
Source cited by the article: anthropic.com/claude-mythos-preview-system-card
Couldn’t checkBenchmark result
GPT-6 Astra solved 85.4% of OpenAI's version of SEC-bench Pro, and GPT-6 Luna 51.75%.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “GPT-6 Astra solved 85.4% of OpenAI’s version, and GPT-6 Luna 51.75%”
The source says No passage could be compared.
Source cited by the article: deploymentsafety.openai.com/gpt-6-luna-october-28-2026-addendum/sec-bench-pro
Couldn’t checkBenchmark result
Claude Opus 5.5 reached code execution in 73% of ExploitBench runs.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “By September, Claude Opus 5.5 reached code execution in 73% of ExploitBench runs”
The source says No passage could be compared.
Source cited by the article: assets.claude.com/fc1b44717c85dc068bc6ba5024219938094694bd.pdf
Couldn’t checkNumber
In the ExploitGym paper, the best model's 157 exploits fell to 45 of 898 with defenses on.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “In the ExploitGym paper, the best model’s 157 exploits fell to 45 of 898 with defenses on”
The source says No passage could be compared.
Source cited by the article: arxiv.org/abs/2605.11086
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 9 Nov 2026, 05:46 UTC. It played no part in the ranking.