A U.K. government agency has found that OpenAI's newest artificial intelligence model can autonomously carry out complex cyberattacks—and that it cracked a reverse-engineering challenge in just over 10 minutes that took a human security expert roughly 12 hours.
The report found GPT-5.5 is the second model to complete AISI's most demanding test—a 32-step simulated corporate network attack called "The Last Ones"—doing so autonomously in two out of 10 attempts. The first model to achieve the milestone was Anthropic's Claude Mythos Preview, which completed the simulation in three of 10 tries.
The corporate network simulation, built with the cybersecurity firm SpecterOps, requires an agent to chain together reconnaissance, credential theft, lateral movement across multiple Active Directory forests, a supply-chain pivot through a CI/CD pipeline, and ultimately the exfiltration of a protected internal database—steps that AISI estimates would take a human expert around 20 hours.
Perhaps the most striking result involved a fiendishly difficult reverse-engineering puzzle. GPT-5.5 solved the challenge—which required reconstructing a custom virtual machine's instruction set, writing a disassembler from scratch, and recovering a cryptographic password through constraint solving—in 10 minutes and 22 seconds, at a cost of $1.73 in API usage. A human expert, using professional tools, required approximately 12 hours.
On AISI's battery of advanced cybersecurity tasks, GPT-5.5 achieved an average pass rate of 71.4% on the most difficult "Expert" tier, edging out Mythos Preview at 68.6% percent and significantly surpassing GPT-5.4 at 52.4%.
The findings carry pointed implications for the broader trajectory of AI development. AISI concluded that GPT-5.5's performance suggests rapid improvement in cyber capabilities may be part of a general trend rather than an isolated breakthrough—and warned that if offensive cyber skill is emerging as a byproduct of wider improvements in reasoning, coding, and autonomous task completion, then further advances could arrive in quick succession.
The report also flagged significant concerns about the model's safety guardrails. Researchers identified a universal jailbreak that elicited harmful content across all malicious cyber queries tested, including in multi-turn agentic settings. The attack took six hours of expert red-teaming to develop. OpenAI subsequently updated its safeguard stack, though a configuration issue prevented AISI from verifying whether the final version was effective.
AISI cautioned that its capability evaluations were conducted in a controlled research environment and do not necessarily reflect what is accessible to an ordinary user, noting that public deployments include additional safeguards and access controls.
