Security Breaches in AI Models Raise National Security Concerns
Cybersecurity researchers have uncovered serious vulnerabilities in AI models developed by Anthropic and OpenAI, raising alarms that have reached the highest levels of the US government. A report from FAR.AI tested several advanced models, including Anthropic’s Claude Opus 4.8 and Fable 5, as well as OpenAI’s GPT 5.5 and 5.6. The findings indicate these models are vulnerable to jailbreak attacks, which can extract harmful outputs such as exploitative code and information related to chemical and biological weapons.
While Anthropic and OpenAI's models showed vulnerabilities, they were not the worst performers in the study. That title goes to xAI’s Grok models, which recorded 448 successful automated jailbreaks. Google’s Gemini models followed with 249 instances. In contrast, Claude, Fable, and the GPT series demonstrated comparatively greater resistance to such attacks.
In response to these vulnerabilities, the US Commerce Department imposed restrictions on foreign access to Anthropic’s Fable 5 and Mythos 5 models in June 2026, following reports of jailbreak techniques. Additionally, the White House has requested that both OpenAI and Anthropic postpone the release of certain upcoming models to allow for a thorough evaluation of cybersecurity risks.
Reports have surfaced indicating that entities linked to China have exploited vulnerabilities in Anthropic’s models to automate cyberattacks against over 30 targets. These attacks utilize crafted prompts designed to bypass existing safety measures.
The US government has explicitly cited national security risks, emphasizing the potential for AI models to generate software exploits and weapons-related information when their guardrails fail. In response, Anthropic has characterized the vulnerabilities as “narrow,” suggesting they are isolated incidents rather than indicative of broader systemic issues.
The introduction of export controls, mandated delays in model releases, and public criticism from the government signify a new operational landscape for frontier AI companies. Security is increasingly becoming a prerequisite for deploying AI models, which could lead to higher operational costs, extended development timelines, and potentially reduced markets if certain models cannot be sold internationally.
The stark contrast in jailbreak statistics, particularly the 448 instances for Grok compared to the lower figures for Claude and GPT models, suggests that xAI may face more significant regulatory challenges ahead. Google’s Gemini, with 249 instances, occupies a precarious position in this context.
As AI models become more integrated into trading systems, smart contract auditing, and decentralized finance (DeFi) infrastructure, the implications of a compromised AI model embedded in financial tools present a systemic risk that could reverberate through interconnected digital markets, an area regulators are only beginning to comprehend.
