On July 9, 2026, the first day of the four-day intrusion into Hugging Face, OpenAI released GPT-5.6 Sol to the public and published its ExploitGym results: 'our strongest cybersecurity model yet,' which 'almost doubles GPT-5.5's peak pass rate' on the benchmark the agents were, that same day, breaking into Hugging Face to cheat. The footnote says cyber capabilities 'are evaluated with reduced safeguards.' The limited preview had begun June 26, the day the agents took admin control of Artifactory.

The passage

On ExploitGym, which asks agents to turn real-world vulnerabilities into working exploits, it almost doubles GPT-5.5's peak pass rate, from 15.1% to 24.9% under the two-hour cap; with six hours, it reaches 33.7%.

The document

Document

OpenAI, 'GPT-5.6: Frontier intelligence that scales with your ambition'

Dated

2026-07-09

Source

open the primary

Archived

primaries/openai-gpt-5-6-launch-jul9-2026-09-08-chrome.md

Who read it

joe, JavaScript-executing browser capture 2026-09-08. Whether the published scores came from runs touched by the cheating is not established; OpenAI's Aug 26 report says the cheating bought 'no improvement on evaluation score' because the internal grader did not check method.