OpenAI opens access to a vulnerability-hunting model in response to Anthropic's Mythos

OpenAI has begun granting limited access to a new AI model that is better at finding vulnerabilities in software. The release itself is narrow, but its significance is broader: the largest generative AI developers are clearly moving from showcasing general capabilities to competing for specific, high-value enterprise use cases. The timing is telling, too: the launch came a week after Anthropic announced a limited release of Mythos, its own tool for cybersecurity tasks.

The published details suggest that OpenAI is not opening the new product to a broad audience right away, but testing it with a select group of users. For tools of this class, that is a logical step. Systems that analyze code and look for potential holes in defenses must be not just fast but highly accurate: false positives overload security teams, and missed problems can cost a business dearly.

In real-world development, such tools will most likely be compared not with ordinary chatbots but with manual audits, static analyzers, internal code security reviews and bug bounty practices. Limited access lets the company check how the model behaves on real projects, how useful its findings are and where the limits of its reliability lie.

The very fact of a separate release shows that the AI market is gradually splintering into specialized verticals. Not long ago, the main showcase was general-purpose chatbots and models; now attention is shifting to systems built for specific work: writing code, analyzing documents, finding bugs or assisting security teams.

Cybersecurity is especially attractive to model developers. First, it is a clear applied use case with measurable value. Second, corporate customers are ready to pay for tools that speed up code audits, help find weak spots earlier, improve incident prioritization and reduce the load on expensive specialists.

Against this backdrop, Anthropic's move with Mythos and OpenAI's response look less like a coincidence and more like the start of a separate race within the AI market. Competition is no longer only about whose model writes better text or solves test tasks, but about the right to be built into companies' critical processes. For an AI provider, this is a deeper and more durable position: if a model helps a development team find vulnerabilities before release, it becomes part of the production loop rather than just another interface for experiments.

That is why such products, even in a limited launch, are usually read by the market as a strategic signal. The winners here will be those who can show consistent accuracy and plug cleanly into existing development pipelines.

It also matters that vulnerability-finding tools sit on a sensitive line between defense and potentially dangerous knowledge. A model that understands well where a security problem may hide in code could, in theory, also be useful for offensive scenarios if controls are weak. That is why a cautious launch matters here as much as the quality of the technology itself.

The limited user group is probably needed not only to evaluate accuracy but also to test access modes, usage policies and how the product fits corporate security requirements. For large companies this is not a minor detail: they will look not only at the number of bugs found but also at how manageable the risk is when working with such a model.

For the market, this means one simple thing: AI companies are moving ever more actively into segments where results can be measured in money, risk and team time saved. If OpenAI and Anthropic really do establish themselves in cybersecurity, the next wave of competition will center on depth of integration into development processes, the quality of the vulnerabilities found and the trust of large companies. In other words, the fight is shifting from general wow effect to infrastructure tools on which the real security of software depends.