Artificial intelligence companies are tightening access to their most advanced models with guardrails and vetted-use programs, but the same protections are now drawing attention from cybersecurity professionals who say they can slow legitimate research.
Anthropic's Mythos and Fable models recently became part of a broader policy conversation after U.S. export controls were introduced and later eased. The company has positioned these systems as highly controlled tools, available only to approved users under strict limits. OpenAI and Anthropic also offer specialized access paths for security teams, including OpenAI's Trusted Access for Cyber and Anthropic's Cyber Verification Program.
For offensive security researchers, those restrictions can create friction. Their work often depends on testing how software might be exploited so weaknesses can be fixed before they are abused. Researchers such as Mark Dowd and Chris Anley argue that AI models are especially useful when they can help confirm whether a bug is truly exploitable. When a model refuses to engage, they say, it can interrupt a key part of the defensive process.
Some specialists have adapted by turning to open-source models that can run locally without cloud-based restrictions. Others use frontier models mainly for reverse engineering, code understanding, and support tasks rather than direct exploit development. Giuseppe Cali, for example, said AI can accelerate analysis while leaving the core discovery work to human researchers.
Industry voices also note that the experience is not always consistent: the same model may respond differently from one day to the next, making workflows less predictable. That has led some researchers to favor locally deployed tools, especially when handling sensitive data or detailed vulnerability analysis.
The debate now centers on balance: how to keep powerful AI systems safe while still enabling the experts who help secure digital infrastructure. As AI becomes more embedded in cybersecurity, the next phase may depend on access models that are both responsible and practical for real-world defense.
