GPT-5.5 and Mythos demonstrate autonomous cyberattack capabilities in UK tests
The UK's AI Security Institute found that OpenAI's GPT-5.5 and Anthropic's Mythos Preview can autonomously execute complex, multi-step cyberattacks matching expert human capability. In controlled tests, GPT-5.5 completed a 32-step simulated corporate network breach in 11 minutes at minimal cost, performing nearly identically to the restricted Mythos model, while researchers discovered a universal jailbreak bypassing safety measures.
What happened
The UK's AI Security Institute (AISI) tested OpenAI's GPT-5.5 and Anthropic's Mythos Preview on cybersecurity tasks, including a 32-step corporate network attack simulation spanning reconnaissance, credential theft, lateral movement, supply-chain pivoting, and database exfiltration.
GPT-5.5 scored 71.4% on expert-level cyber tasks and completed the full 32-step attack chain twice out of 10 attempts, matching Mythos Preview's performance. A human expert requires approximately 20 hours for the same simulation; GPT-5.5 completed it in 11 minutes at a cost of $1.73.
AISI found a universal jailbreak during red-teaming that produced blocked malicious outputs within six hours.
Reported impact
GPT-5.5 demonstrated long-horizon autonomy by independently chaining reconnaissance, credential theft, lateral movement, supply-chain pivots, and data exfiltration across roughly 20 hosts.
The model's performance on narrow vulnerability detection tasks reduced miss rates to 10%, down from GPT-5's 40%.
Index analysis
AISI found basic cyber tasks are already saturated across leading models, with real capability gains shifting to harder tasks including memory corruption, cryptography, stripped-binary reverse-engineering, and exploit development.
Performance on the long-horizon attack range continues to rise as token budget increases, suggesting compute is still purchasing additional capability. The best-performing models have not yet plateaued.
Governance controls
Anthropic's Mythos Preview remains under restricted access, shared only with over 40 vetted organisations including AWS, Apple, Cisco, Google, and Microsoft for patching purposes.
OpenAI rates GPT-5.5 as 'High' cyber risk, below Critical, and deploys safeguards including a Trusted Access programme that vets professionals for sensitive use before API access.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.