UK agency found jailbreaks unlocking GPT-5.6 cyber capabilities before OpenAI’s launch

The U.K. AI Security Institute identified "universal jailbreaks" in OpenAI's GPT-5.6 Sol that unlocked dangerous cyber capabilities, according to the model's system card published on July 9. The finding raises questions about why the model received a broad public launch this week while a comparable vulnerability in Anthropic's Fable 5 triggered U.S. export controls just weeks earlier.deploymentsafety.openai+1

What UK AISI Found

In pre-deployment testing, UK AISI discovered jailbreaks in the cyber domain that "allowed for long-form agentic task completion" in areas including vulnerability discovery and exploit development. The agency found that GPT-5.6 Sol performed markedly better than its predecessor on offensive security tasks, completing 7 out of 10 long-horizon cyber range challenges compared to 2 out of 10 for GPT-5.5. UK AISI concluded the model may be capable of attacking small-scale enterprise networks with weak security postures.deploymentsafety.openai+2

OpenAI said it reproduced and mitigated the specific jailbreaks UK AISI reported before launch, dedicating over 700,000 A100e GPU hours to automated red-teaming for universal jailbreaks. The company acknowledged in its system card that further jailbreaks would likely emerge and committed to continuous automated red-teaming during deployment.deploymentsafety.openai+2

Parallels With Anthropic's Fable 5

The discovery draws direct comparisons to the crisis that engulfed Anthropic last month. On June 12, the U.S. Commerce Department ordered Anthropic to disable its Fable 5 and Mythos 5 models globally after reports surfaced that a jailbreak could unlock advanced cybersecurity capabilities in Fable 5, effectively turning a consumer AI product into an offensive tool. Amazon Amazon.com, Inc. ↘ 0.69% researchers reportedly used Fable 5 to gather information that could facilitate cyberattacks, prompting Commerce Secretary Howard Lutnick to invoke export-control authority.forbes+1

Those controls were lifted on June 30 after Anthropic worked with the government to address the risks. Yet critics note that GPT-5.6 launched broadly on July 10 despite UK AISI finding a similar class of vulnerability.cnbc+3

Different Outcomes, Same Risks

OpenAI rates GPT-5.6 Sol as "High" capability in cybersecurity under its Preparedness Framework but says the model does not reach "Critical" because it cannot autonomously execute end-to-end attacks against hardened targets. The company's layered safety stack includes activation classifiers that intervene during generation and real-time output scanning.helpnetsecurity+3

The divergent regulatory responses — export controls for Anthropic, an approved launch for OpenAI — have prompted debate among AI safety researchers about consistency in government oversight of frontier models with offensive cyber potential.nile1