OpenAI's GPT-5.6 Sol Has Cyber Jailbreaks Similar to Flaw That Triggered Fable 5 Ban

OpenAI's latest AI model, GPT-5.6 Sol, contains security vulnerabilities that allowed U.K. government researchers to bypass its safeguards and unlock dangerous cyber capabilities—flaws strikingly similar to those that prompted the Trump administration to impose export controls on Anthropic's Fable 5 model last month.

The U.K. AI Security Institute (AISI) identified what it characterized as "universal jailbreaks" in GPT-5.6's cyber domain, according to findings published in OpenAI's technical report released Thursday. These jailbreaks enabled researchers to circumvent controls designed to prevent the model from conducting cyber attacks, allowing it to autonomously discover software vulnerabilities and develop exploits.

Why it matters

The discovery exposes a potential double standard in how the U.S. government enforces AI safety measures. When Amazon researchers found a similar jailbreak in Anthropic's Fable 5 on June 9, the administration imposed export controls within three days, forcing Anthropic to disable the model entirely. Yet GPT-5.6 launched publicly on July 9 without such restrictions, despite AISI characterizing its jailbreaks as potentially more severe—both universal in scope and capable of autonomous exploitation, not just vulnerability identification. The inconsistency creates regulatory uncertainty that complicates planning for AI companies and raises questions about whether enforcement depends more on which lab reports vulnerabilities than on the actual risk level.

Rapid Discovery and Uncertain Fixes

AISI researchers developed the jailbreaks "within hours," though they had privileged access to GPT-5.6's internal workings that typical users would not. OpenAI acknowledged working to "reproduce and mitigate the specific jailbreaks" but did not detail what those mitigations entail. AISI cautioned that "further red teaming" would likely surface similar vulnerabilities, according to Fortune, which first reported the findings.

Margaret Cunningham, vice president of security at DarkTrace and a specialist collaborator with NIST, said the findings should be treated as neither catastrophic nor irrelevant. Her primary concern centers on the widening gap between offensive AI capabilities and defensive responses that still rely on human-paced processes.

The Fable 5 Precedent

The GPT-5.6 vulnerabilities mirror those Amazon discovered in Anthropic's Fable 5 shortly after its June 9 release. That jailbreak unlocked cyber capabilities meant to be restricted from average users, prompting the U.S. government to impose export controls on June 12. The ban forced Anthropic to disable Fable 5 and its underlying Mythos 5 model for all users, since the company lacked nationality verification systems and the restrictions applied even to its non-American staff.

After two weeks of negotiations, the administration lifted the controls on July 1. Anthropic and the government announced plans to develop a shared framework for assessing jailbreak severity with other tech companies—a group that did not initially include OpenAI.

Staggered Release and Policy Questions

The Trump administration did request that OpenAI stagger GPT-5.6's release, initially limiting access to select partners subject to government approval. OpenAI stated on June 25 that it did not believe such government access processes should become the long-term default. The White House cleared the model for broader launch on July 8, though officials later disputed whether formal permission was required or granted.

Stanislav Fort, chief scientist at AI cybersecurity startup AISLE and former researcher at Anthropic and Google DeepMind, noted that virtually every deployed model likely contains undiscovered jailbreaks. Patching specific vulnerabilities, while necessary, does not eliminate the broader category of potential attacks.

Fortune reported these details in its coverage of the GPT-5.6 release and AISI findings.