Bottom line: While China seeks access to US cyber AI models, the US industry is racing to deploy these models for defensive measures quickly enough – but time is running short.
New generative AI models such as Anthropic Claude Mythos and OpenAI GPT 5.5-Cyber identify software vulnerabilities and can execute cyberattacks – faster than regulation or defensive measures can be built up. US experts estimate that China could gain access to comparable models or develop its own cyber AI weapons within 6 to 12 months.
Claude Mythos (available since April 2026) and GPT 5.5-Cyber demonstrate remarkable capability for automated identification of security gaps and execution of attack scenarios. The US administration is in a race: on one hand, American technology companies should be able to develop and test defensive tools against AI-driven cyberattacks; on the other hand, China is suspected of copying US technologies through distillation attacks – attackers train their own “student” models using the outputs of an advanced “teacher” model.
Rob T. Lee from the SANS Institute describes the situation as a “hurricane warning, not a wall of protection”. The new models compress attack preparation that previously required days, weeks or years into seconds. This fundamentally shifts the balance of power between attackers and defenders: defenders must expect a manifold higher volume of attacks.
Anthropic and OpenAI initially restricted access to their new models to small groups of trusted security experts. In April, Anthropic declined a request from China to trial Mythos. Meanwhile, Anthropic is intensifying its rollout: the company announced that “Mythos-class” models will be made available to all customers in the coming weeks. This week, Anthropic shared Mythos with approximately 150 new organizations across 15 countries, with each organization required to meet unspecified security requirements.
Lee Klarich from Palo Alto Networks describes the dilemma: “There is a balance point between limited and broad availability in order to find and remediate as many critical vulnerabilities as possible before attackers gain access to these models.” The Trump administration has pursued a cautious approach to regulating frontier models so far, while Anthropic warned in a blog post that AI will soon be powerful enough to suppress citizens on an unprecedented scale or even shift the balance of power between nations.
Source: www.politico.com · Published June 7, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.6.5.