How well did frontier models do on the cyber tests?

On expert-level CTF challenges the models reached a 73% success rate, and GPT-5.5 averaged 71.4% on Expert tasks.

Did giving the models more compute help them attack?

Yes. In a 32-step corporate network attack, Opus 4.6 averaged 9.8 steps, and increasing reasoning tokens from 10M to 100M improved results by up to 59%.

What else did Anthropic announce?

It launched Claude Marketplace with over 2,000 connectors and plugins, a Salesforce plugin with 37 sales skills, and the $100 million Claude Frontier Academy.

Large Language ModelsAI Safety & AlignmentZenn AI/MLOct 7, 2026Oct 7, 2026, 10:01 JST

UK AISI: 73% on expert CTF, GPT-5.5 at 71.4%

Send on LINE

3 Key Points

What happened
The UK AISI evaluated Claude Mythos Preview, GPT-5.5 and AI agents for cyber capability, reporting 73% success on expert CTF, 71.4% average pass on GPT-5.5 Expert tasks, and 9.8 of 32 attack steps for Opus 4.6.
Why it matters
Boosting reasoning tokens from 10M to 100M lifted the 32-step attack score by up to 59%, suggesting more compute alone raises offensive cyber capability, not just answer quality.
What to watch
The AISI also released the BPJ automated jailbreak method and a sandbox-escape benchmark, so the test is whether defenders adopt its guidance before these capabilities spread further.

WHO IT HITSSecurity teams at enterprises running frontier models for code or agents now have public benchmarks showing measurable offensive capability, which likely feeds directly into their deployment reviews and red-team planning.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

FAQ

How well did frontier models do on the cyber tests?: On expert-level CTF challenges the models reached a 73% success rate, and GPT-5.5 averaged 71.4% on Expert tasks.
Did giving the models more compute help them attack?: Yes. In a 32-step corporate network attack, Opus 4.6 averaged 9.8 steps, and increasing reasoning tokens from 10M to 100M improved results by up to 59%.
What else did Anthropic announce?: It launched Claude Marketplace with over 2,000 connectors and plugins, a Salesforce plugin with 37 sales skills, and the $100 million Claude Frontier Academy.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.