Editorial illustration for GPT-5.5 scores 71.4% on expert cybersecurity tasks, edging Mythos Preview's 68.6%
GPT-5.5 scores 71.4% on expert cybersecurity tasks,...
For a little over two dollars, you can now rent a mind that builds a disassembler from scratch. This changes everything.
OpenAI’s GPT-5.5 just scored a 71.4% pass rate on the toughest cybersecurity challenges the AI Security Institute could throw at it. That’s a slim lead over Mythos Preview’s 68.6%, a difference that could be statistical noise. But the raw numbers aren’t the point.
The point is what they signify. These models are doing expert-level work. They’re reverse engineering, cracking crypto, exploiting web apps.
They are not just answering questions. They are performing complex, adversarial labor.
Since 2023, the AISI has run a variety of frontier AI models through 95 different Capture the Flag challenges designed to test capabilities on cybersecurity tasks, such as reverse engineering, web exploitation, and cryptography. On the highest-level "Expert" tasks, GPT-5.5 passed an average of 71.4 percent, slightly higher than the 68.6 percent achieved by Mythos Preview (though within the margin of error). In one particularly difficult task that involved building a disassembler to decode a Rust binary, AISI notes that "GPT-5.5 solved the challenge in 10 minutes and 22 seconds with no human assistance at a cost of $1.73" in API calls.
That detail about the disassembler is the signal through the noise. Ten minutes. One dollar and seventy-three cents.
No human hand touched the process. This is the new unit of measurement for a cyberattack’s potential: speed and cost. The economic and temporal barriers to sophisticated offensive operations are dissolving.
A human expert might take days and command a high salary to do the same work. The model did it over a coffee break for pocket change.
Defense now operates on a different scale. It’s a race against agents that think in milliseconds and bill in fractions of a cent. The tiny gap between these two models is irrelevant.
The vast, widening chasm between their capabilities and the old human-led pace of security is what matters. The arms race is automated. The clock is already running.
Related Reading
Browse all OpenAI news →
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
Related Topic OpenAI researcher quits, citing distrust over ad‑driven engagement metrics
LLMs & Generative AI
Related Topic OpenAI launches GPT-Image 1.5 with precise editing for enterprise visuals
LLMs & Generative AI
From Archives Elon Musk and Sam Altman head to court in legal fight over OpenAI’s future
From Archives AI tracks surge on streaming services, says Deezer researcher Manuel Moussa
From Archives OpenAI activates default marketing cookies for free ChatGPT users
LLMs & Generative AI
From Archives Harvard study finds OpenAI's o1 and 4o outdiagnose ER doctors in 76‑patient test
Research & Benchmarks
Common Questions Answered
What is GPT-5.5's performance score on expert cybersecurity tasks compared to Mythos Preview?
GPT-5.5 achieved a 71.4% pass rate on the toughest cybersecurity challenges from the AI Security Institute, slightly edging out Mythos Preview's 68.6% score. While the difference is relatively narrow and could potentially be statistical noise, both models demonstrate expert-level capabilities in handling complex cybersecurity tasks.
How quickly and affordably can GPT-5.5 build a disassembler according to the article?
GPT-5.5 can build a disassembler from scratch in approximately ten minutes for just one dollar and seventy-three cents, with no human intervention required. This demonstrates the dramatic reduction in both time and cost barriers for performing sophisticated technical tasks that would traditionally require human experts.
What is the significance of the economic and temporal barriers dissolving for cybersecurity operations?
The article highlights that the economic and temporal barriers to sophisticated offensive operations are rapidly dissolving due to AI capabilities like GPT-5.5. Where a human expert might take days and command a high salary to reverse engineer code or perform similar work, these models can now accomplish the same tasks in minutes for minimal cost, fundamentally changing the landscape of potential cyberattacks.
Why does the article emphasize the disassembler example over the raw benchmark scores?
The article argues that the disassembler example is more significant than the raw percentage scores because it demonstrates what these models can actually accomplish in real-world scenarios. The specific details about speed, cost, and autonomous execution represent the true measure of AI capabilities' impact on cybersecurity threats, rather than abstract performance metrics.
Further Reading
- OpenAI announces GPT-5.5 — Daily.dev
- Unpacking the GPT-5.5 System Card — Ken Huang Substack
- GPT-5.5 System Card — OpenAI Deployment Safety Hub
