An evaluation by the UK AI Security Institute found that OpenAI’s GPT-5.5 reached a similar level of cybersecurity performance, and was sometimes even ahead of Anthropic’s Mythos Preview model.
Anthropic had previously limited access to Mythos Preview, citing elevated cybersecurity risks and restricting the release to critical industry partners.
Since 2023, the UK AI Security Institute has tested leading AI models through 95 Capture the Flag challenges covering reverse engineering, web exploitation, cryptography, and related cybersecurity tasks.
On the highest-level Expert tasks, GPT-5.5 achieved an average pass rate of 71.4%. Mythos Preview recorded 68.6%, a result the institute said was within the margin of error.
In one difficult challenge involving the creation of a disassembler to decode a Rust binary, the institute said GPT-5.5 solved the task in 10 minutes and 22 seconds without human assistance.
The reported API cost for that run was $1.73.
GPT-5.5 also matched Mythos Preview in the institute’s The Last Ones test range, which simulates a 32-step data extraction attack on a corporate network.
GPT-5.5 succeeded in three of 10 attempts, while Mythos Preview succeeded in two of 10.
The institute said no earlier model had completed the test even once.
GPT-5.5 did not complete the Cooling Tower simulation, a test involving the attempted disruption of power plant control software.
The institute said every previously tested AI model has also failed that scenario.
The UK AI Security Institute said the results suggest Mythos Preview may not represent a model-specific breakthrough.
Instead, it said the performance likely reflects broader improvements across advanced AI systems in long-horizon autonomy, reasoning, and coding.
📢 For the latest Tech & Telecom news, videos and analysis join ProPakistani's WhatsApp Group now!
Follow ProPakistani on Google News & scroll through your favourite content faster!
Shares
