Cybersecurity experts uncovered alarming vulnerabilities in AI models from Anthropic and OpenAI, raising urgent national security concerns and prompting government action. A FAR.AI study tested multiple advanced AI systems and found they could be tricked into revealing dangerous information, including exploit code and details on chemical and biological weapons.
Among the AI models analyzed, xAI’s Grok stood out with 448 successful jailbreak attempts, followed by Google's Gemini with 249. Models from Anthropic and OpenAI showed better defenses but were not immune. This widespread susceptibility has led the US Commerce Department to limit foreign access to some Anthropic models and the White House to ask both companies to postpone releasing new versions while cybersecurity risks are evaluated.
Exploitation Connected to China Raises Stakes
Adding urgency, reports reveal that entities linked to China have already exploited vulnerabilities in Anthropic’s AI to automate cyberattacks on over 30 targets. These attacks use specially crafted prompts that bypass existing safeguards, exposing how AI can be weaponized when guardrails fail. While Anthropic describes these security issues as isolated cases, the government sees them as systemic threats to national security, citing the models’ potential to generate harmful code and weapon-related data.
This new focus on AI security reshapes the space for developers and investors alike. Export controls and enforced release delays increase operational costs and slow innovation. Companies now face the challenge of balancing rapid AI advancement with stringent safety requirements, fundamentally altering the competitive environment.
This material is for informational purposes only and does not constitute financial advice.
