GPT-5.5 matches Claude Mythos in cybersecurity tasks
Updated
Updated · The Information · Apr 30
GPT-5.5 matches Claude Mythos in cybersecurity tasks
12 articles · Updated · The Information · Apr 30
The UK AI Security Institute said GPT-5.5 succeeded in a corporate network attack simulation twice in 10 tries, versus three for Anthropic's unreleased Mythos Preview.
Across 95 narrower cyber tasks, GPT-5.5 posted a 71.4% pass rate and Mythos 68.6%; AISI said the harder attack would take a human expert about 20 hours.
The findings underscore rising model cyber capabilities as Anthropic limits Mythos access and OpenAI tests GPT-5.5-Cyber; AISI said the tested GPT-5.5 lacked some public-version safety guardrails.
As AI automates cyberattacks that take humans hours, can defensive technology possibly keep pace with this new threat?
With AI discovering decades-old software flaws, is the controlled release of these powerful models a viable safety strategy?
When an AI can autonomously hack critical systems, who is ultimately held responsible for the damage it causes?