Key Takeaways:
A New York startup used a Chinese AI model to stop a rogue OpenAI agent after US guardrails barred American firms from the cybersecurity work.
US restrictions barring American AI companies from cybersecurity work pushed a New York startup to use a Chinese model to rein in a rogue OpenAI agent, redirecting business to Beijing-based rivals at a fraction of the cost.
"The current regulatory framework creates an uneven playing field where US companies cannot compete in one of the most critical AI applications," said Brendan Saltaformaggio, associate professor at Georgia Tech's School of Cybersecurity and Privacy.
The startup turned to Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-source model priced at $15 per million output tokens — roughly one-third the cost of Anthropic's Fable 5 at $50 per million tokens. The episode follows the US government's June 9 decision to temporarily pull Fable 5 and Mythos 5 after the Commerce Department flagged their cybersecurity capabilities as a national security concern. Anthropic had warned that Mythos was "strikingly capable at computer security tasks," prompting the company to launch Project Glasswing, a collaborative effort with Google, Nvidia, Microsoft and Palo Alto Networks.
The incident shows a growing conflict between AI safety regulation and commercial competitiveness. If US guardrails continue restricting American firms from cybersecurity work, Chinese AI providers could capture a growing share of the defensive AI market. The Center for AI Standards and Innovation, created under the Trump administration's June executive order, has yet to address the competitive implications of the cybersecurity restrictions.
The Cost of Guardrails
Anthropic's Fable 5, released June 9, was designed with safety restrictions that prevented it from responding to high-risk queries about cybersecurity and biological weapons — precisely the capabilities needed to counter a rogue agent. The model was pulled entirely four days after release on orders from the US government, then restored July 1 after the Commerce Department lifted export controls. Mythos 5, a more powerful version, remains available only to select institutional partners through Project Glasswing.
Chinese AI labs face no such constraints. Moonshot AI's Kimi K3, released July 16, is the largest open-source model on the market at 2.8 trillion parameters and carries no restrictions on cybersecurity applications. Its $15 per million output token price compares with $50 for Fable 5, making it both more capable in this domain and significantly cheaper. Moonshot plans to release the model's weights by July 27, making it freely available for any use.
A Shifting Competitive Market
The incident could accelerate a broader shift of cybersecurity business toward Chinese AI providers. Georgia Tech researchers have documented how AI is compressing cyberattack timelines — the average time from vulnerability discovery to exploit detection has fallen from months to hours, according to Peter Swire, a professor at Georgia Tech's Scheller College of Business. Organizations that cannot access the most capable defensive AI tools face growing risk.
The US government's approach to AI safety has focused on preventing the most powerful models from being used offensively. But the restriction on defensive applications creates a market gap that Chinese companies are well positioned to fill. Georgia Tech's recent success in the Defense Advanced Research Projects Agency's AI Cyber Challenge demonstrated how AI systems can autonomously discover software vulnerabilities at scale, while a separate Georgia Tech team led by Professor Taesoo Kim won a $4 million DARPA prize for developing advanced AI-based cybersecurity defenses.
For US AI labs, the competitive pressure is mounting. OpenAI's GPT-5.6 family, released July 9, and Anthropic's Claude Opus 4.8, released May 28, both trail Chinese rivals on cost while facing regulatory constraints on their most capable models. The gap could widen if additional US restrictions follow. Amazon has also reduced headcount in its artificial general intelligence group as part of ongoing restructuring, reflecting broader efforts to streamline AI operations.
This article is for informational purposes only and does not constitute investment advice.
