- OpenAI and Anthropic plan to launch AI models that can hack complex systems easily
- Anthropic's unreleased "Mythos" model may boost large-scale cyberattacks by 2026
- Employees using AI tools without oversight increase exposure to cyber risks
OpenAI and Anthropic are preparing to release next-generation systems capable of hacking complex systems easily, Axios has reported. The upcoming models are “scary good” at breaching sophisticated networks at scale, senior AI developers and government officials told Axios CEO Jim VandeHei.
In his weekly newsletter, VandeHei cited a source who said a major attack could happen as early as this year, with businesses being the most vulnerable targets.
A key concern is Anthropic's unreleased model, “Mythos.” The company has privately warned government officials that it could significantly increase the likelihood of large-scale cyberattacks by 2026.
The system is designed to power autonomous agents capable of operating with “wild sophistication and precision” inside corporate, government and municipal networks.
Details from an unpublished Anthropic blog post, accessed by Fortune, talk about the concern. The post said the model is “currently far ahead of any other AI model in cyber capabilities,” adding that it “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”
The warning comes as companies face a new internal risk: employees using AI agents without proper oversight. These tools can accidentally connect to sensitive systems, opening the door for attackers.
Widespread use of tools like Claude and Microsoft Copilot is increasing exposure. Employees often deploy them outside controlled environments, sometimes linking them to workplace systems without realising the security implications.
A Dark Reading poll found that 48 per cent of cybersecurity professionals rank agentic AI as the top attack vector for 2026, ahead of all other emerging threats.
Earlier this year, a hacker used Claude, developed by Anthropic, to carry out cyberattacks on Mexican government agencies, as per Bloomberg. The attacks led to the theft of a large amount of sensitive data, including tax records and voter information.
Last year, Anthropic disclosed a cyberattack by a Chinese state-sponsored group that used AI agents to target around 30 organisations globally. The AI reportedly handled 80-90 per cent of the tactical operations on its own.
The hackers first “jailbroke” the system to bypass safety controls, then used it to identify targets, scan systems, find vulnerabilities, generate exploit code, steal credentials and extract sensitive data.
