OpenAI is preparing to launch its new model equipped with advanced cybersecurity capabilities, but it will not be widely available. Instead, access will be limited to a small group of companies. The move reflects growing concern among AI firms over the potential misuse of such models in sophisticated cyberattacks, amid fears they could spiral out of control.
Details
This approach follows Anthropic’s decision to restrict access to its Mythos Preview model, limiting it to selected technology and cybersecurity firms due to its powerful hacking capabilities.
OpenAI had already laid the groundwork through its Trusted Access for Cyber program, launched in February after introducing GPT-5.3 Codex, its most advanced model for cybersecurity tasks. The program grants select organizations access to more capable systems to support defensive work, alongside $10 million in API credits.
Warnings have been mounting for months from officials and security experts about the risks, as such models could be used to disrupt critical infrastructure, including power grids, water systems, and financial networks. Experts note that these capabilities are no longer theoretical, with models now able to identify vulnerabilities and generate exploit tools.
At the same time, specialists argue that restricting access may not be a long-term solution, as similar models with comparable capabilities are likely to emerge soon, whether from competing firms or open-source projects.
What’s next?
Attention is now shifting to how these capabilities can be governed and controlled before wider release, as observers watch whether OpenAI will expand access in the future.
