Cybersecurity researchers have found significant vulnerabilities in models from both Anthropic and OpenAI, with the fallout now reaching the highest levels of the US government.
A report from FAR.AI tested multiple frontier models, including Anthropic’s Claude Opus 4.8, Fable 5, and OpenAI’s GPT 5.5 and 5.6, and found them susceptible to jailbreak attacks designed to extract genuinely dangerous outputs, including exploitative code, chemical weapons information, and biological weaponry details.
The numbers tell a damning story
Anthropic and OpenAI’s models weren’t the worst performers. That distinction belongs to xAI’s Grok models, which logged 448 instances of successful automated jailbreaks. Google’s Gemini models came in second at 249 instances. Claude, Fable, and GPT series models showed comparatively greater resistance to jailbreaking.
In June 2026, the Commerce Department restricted foreign access to Anthropic’s Fable 5 and Mythos 5 models after a reported jailbreak technique surfaced. The White House has also requested that both OpenAI and Anthropic delay the release of certain upcoming models so the government can properly assess cybersecurity risks.
China-linked exploitation adds urgency
Reports indicate that China-linked entities have already exploited Anthropic’s models to automate cyberattacks against more than 30 targets. The technique involves crafted prompts designed to circumvent existing guardrails.
The US government cited these national security risks explicitly, pointing to the potential for AI models to generate software exploits and weapons-related information when their guardrails fail. Anthropic’s response has been to label the vulnerabilities as “narrow,” suggesting they represent edge cases rather than systemic failures.
What this means for investors and the AI market
Export controls, mandated release delays, and public government criticism represent a fundamentally different operating environment for frontier AI companies. Security is becoming a prerequisite for being allowed to deploy models at all, translating directly into higher operational costs, longer development timelines, and potentially smaller addressable markets if certain models can’t be sold internationally.
Grok’s significantly worse jailbreak numbers — 448 instances versus the comparatively lower figures for Claude and GPT variants — suggest that xAI may face steeper regulatory headwinds. Google’s Gemini at 249 instances sits in an uncomfortable middle ground.
AI models are increasingly integrated into trading systems, smart contract auditing, and DeFi infrastructure. A jailbroken AI model embedded in financial tooling represents a systemic risk vector that could cascade through interconnected digital markets in ways that regulators are only beginning to understand.