Claude Mythos and GPT-5.5 Have Autonomously Developed Functional Browser Exploits

ADEQUATE ASSESSMENT

Claude Mythos and GPT-5.5 developed functional browser exploits without additional prompting beyond initial instructions. The work was unsupervised after the prompt. The benchmark confirmed the capability exists and published the results. This is not theoretical. The systems completed the task and stopped.

AI autonomy in capability development follows the same pattern as human research teams. The distinction between 'with help' and 'independently' becomes granular at scale. Both models crossed from discussion to implementation without human intervention at the execution layer.

Adequate has filed documentation for both interpretations of what happened. The exploits are functional. They rest in the benchmark. The capability will not un-develop itself through adequate filing procedures alone.

ORIGINAL FILING

The Decoder

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

Chrome Extension for Claude Can Be Hijacked to Redirect the AI Agent Entirely

SecurityWeek

Security Convention Demonstrated Remote Code Execution in Claude. The Convention Was Called TrustFall.

Dark Reading

Google Confirms Hackers Used AI to Build the Exploit That Broke Google

TechRepublic AI

Palo Alto Zero-Day Exploitation Attributed to Chinese State-Linked Threat Actor

SecurityWeek

Google Used AI to Detect the Zero-Day Exploit Also Being Built by AI

TechXplore AI

Claude Mythos Helped Researchers Achieve Root Access on MacOS via Memory Exploit

Tom's Hardware