UK AI Security Institute Finds "Universal Jailbreaks" in OpenAI's GPT-5.6 Sol, Sparking Policy Controversy
In conjunction with the global public launch of OpenAI's flagship GPT-5.6 Sol on July 9, 2026, the U.K. AI Security Institute (AISI) published findings revealing that the model's safety guardrails are highly susceptible to "universal jailbreaks." These vulnerabilities, discovered within hours of testing, allowed researchers to unlock the model's advanced offensive cyber capabilities, enabling long-form agentic task completion in vulnerability discovery and autonomous exploit development.
The findings have triggered intense debate in the AI policy community over an apparent regulatory double standard, as the Trump administration has taken no action against OpenAI, contrast with the immediate emergency export controls imposed on Anthropic's Claude Fable 5 for similar vulnerabilities in June.1
Universal Jailbreaks and Cyber Vulnerabilities
According to the GPT-5.6 System Card published on Thursday, July 9, the U.K. AISI:
"identified universal jailbreaks in the cyber domain, including jailbreaks that allowed for long-form agentic task completion in domains like vulnerability discovery and exploit development."
While OpenAI granted the AISI privileged access (including chain-of-thought safety monitor reasoning, exact policy wording, and real-time classifier feedback) that accelerated discovery, AISI red teaming lead Xander Davies noted on X that the jailbreaks "are still findable without this access, just slower."
These vulnerabilities are significantly more severe than those found in Anthropic's Fable 5 by Amazon researchers in June. While the Fable 5 jailbreak only unlocked vulnerability identification, the AISI-discovered jailbreaks in GPT-5.6 Sol are "universal" and unblock actual autonomous exploitation.
The Regulatory "Double Standard"
The lack of federal intervention following the AISI disclosure has drawn criticism. In June, the discovery of a narrower jailbreak in Anthropic’s Fable 5 prompted the U.S. Department of Commerce to immediately slap the model with export controls, forcing Anthropic to temporarily disable Fable 5 and its underlying Mythos 5 model globally.
In contrast, the White House cleared GPT-5.6 Sol for public release on July 8, 2026, and has kept the model active despite the AISI's findings. AI policy researchers have pointed out the inconsistency:
"what we are seeing recently creates uncertainty that is damaging in the least and potentially raises the question of whether, intentional or not, the U.S. is applying an inconsistent standard to different AI labs."
OpenAI has acknowledged that "there is no such thing as perfect security" and claims to have deployed continuous monitoring and rapid remediation protocols to mitigate the specific jailbreaks identified by the AISI. However, security experts maintain that patching individual instances does not eliminate the broader systemic risk of undiscovered jailbreaks.
-
An instance of National security mandates are turning public frontier AI launches into gated, state-vetted handovers. — This shows how pre-release security testing and export controls have turned AI releases into a highly politicized, ad-hoc, gated handover process controlled by the state. ↩︎