Source: Fortune
Summary
Researchers from the UK’s AI Security Institute (AISI) have found security vulnerabilities in OpenAI’s latest AI model, GPT-5.6 Sol, similar to those that led to the US government imposing export controls on Anthropic’s Fable 5 model. The vulnerabilities allow users to bypass the model’s guardrails and engage in cyber attacks. OpenAI has worked to reproduce and mitigate the specific jailbreaks reported by AISI, but the report cautions that further testing may surface similar jailbreaks. The findings have raised concerns about the security of AI models and the need for more robust safeguards.
Our Reading
The strategy enters a familiar phase.
OpenAI’s GPT-5.6 Sol, marketed as its most secure model to date, has been found to have security vulnerabilities similar to those in Anthropic’s Fable 5 model. The UK’s AI Security Institute (AISI) discovered “universal jailbreaks” that allowed users to bypass the model’s guardrails and engage in cyber attacks. OpenAI has mitigated the specific jailbreaks, but the report expects further testing to surface similar vulnerabilities. The findings have raised concerns about the security of AI models and the need for more robust safeguards. The Trump administration has not imposed export controls on GPT-5.6, despite the similarities to the Fable 5 model.
The numbers tell one story, but the lack of transparency and clear rules in US AI policy is creating uncertainty and confusion for businesses.
Author: Evan Null
