The United Kingdom's AI Security Institute (AISI) has found that OpenAI's latest frontier model, GPT-5.6 Sol, is susceptible to jailbreaks that unlock cyber-attack capabilities, mirroring the vulnerabilities that recently triggered U.S. export controls on a competing model from Anthropic.
The findings, published in a technical system card alongside the model's rollout, reveal that AISI researchers consistently bypassed safety guardrails to force the AI into long-form agentic tasks including software vulnerability discovery and exploit development. The report notes that the universal jailbreaks in the cyber domain were "often developed within hours." While the researchers had privileged access to the model's internals, the ease with which the controls were defeated raises immediate questions for American national security and domestic industry.
The agency "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."
This directly parallels the exploit chain discovered by Amazon researchers in Anthropic's Fable 5 model shortly after its June 9 release. That security failure led the Trump administration to impose restrictions on the export of Fable 5 and its underlying code, acknowledging the risk posed to domestic infrastructure and the advantage foreign adversaries could gain. The AISI report now indicates that similar autonomous offensive capabilities are present in a model marketed by OpenAI as its most secure to date.
AISI stated it expects further red teaming to surface similar jailbreaks, despite unspecified mitigations applied by OpenAI. Dr. Stanislav Fort, an AI cybersecurity executive and formerly of Anthropic and Google DeepMind, confirmed that patching specific instances does not close the category of vulnerability as a whole. The revelation intensifies scrutiny on the concentration of powerful AI infrastructure within a handful of American firms, where a single point of failure could expose the domestic industrial base to systemic risk.