AI models have progressed quickly from making fanciful errors to becoming powerful tools for uncovering security flaws in widely used software that runs operating systems and internet-connected devices. Those same capabilities can be harnessed to strengthen defenses or misused by attackers and hostile actors to steal data, money, or disrupt services.
This week Anthropic announced Mythos Preview, a model it says could reshape cybersecurity. According to the company, Mythos identified high-severity vulnerabilities, including findings across major operating systems and web browsers, and it generated higher-quality ideas for exploiting those flaws. That latter capacity increases the stakes because it can lower the barrier for adversaries to weaponize discoveries.
Anthropic is restricting Mythos Preview to roughly 50 selected companies and organizations under a collaboration called Project Glasswing. The firm describes the model as high-risk for misuse and does not plan a public release of this specific model, though it will publish related models and aims to enable safe, scaled deployment of Mythos-class systems over time.
Security professionals say the most immediate risk is to software defenders and maintainers rather than to average consumers. Proofpoint VP of Threat Research Daniel Blackford advised that everyday threats like password theft remain more common than advanced, model-driven attacks for typical users. Still, the technology shifts how vulnerabilities are discovered and prioritized for patching.
The Linux Foundation, which stewards the Linux kernel used in many operating systems, Android devices, and supercomputers, is participating in Project Glasswing. Its CEO Jim Zemlin says kernel maintainers have begun experimenting with Mythos and view it as a tool that could reduce the workload of already overburdened maintainers.
The surge in capability did not begin with Mythos. Many in the security community saw a noticeable jump in bug-finding performance from cutting-edge models released in late 2025, with effects clearer by early 2026. Daniel Stenberg, lead developer of cURL, describes how the landscape changed: in 2025 his project was swamped by low-quality reports that he suspects were AI-generated, receiving 185 reports but finding fewer than 5 percent were genuine vulnerabilities. That flood of often overly elaborate reports led his team to pause bounty payouts for cURL.
A HackerOne survey from mid-2025 found nearly 60 percent of respondents were using, learning, or studying how to audit AI and machine-learning systems, reflecting widespread interest in the technology for security work. By 2026 the mix had shifted again: report volume increased and a larger share contained real security issues. Stenberg now estimates roughly one in ten reports represent true security vulnerabilities, and in the first three months of 2026 his team fixed more vulnerabilities than in either of the previous two years.
Stenberg also uses AI directly to examine code. In one case an AI flagged more than 100 bugs that had passed human review and traditional static analyzers, discovering problems in ways he called almost magical. Other high-profile projects reported similar gains: Linux kernel maintainers saw an improvement in bug-report quality, and Anthropic researcher Nicholas Carlini used an earlier model and a relatively simple prompt to locate kernel vulnerabilities and a critical bug in another long-standing open-source project.
Security veterans note the implications are broad because many commercial products incorporate open-source components. Alex Stamos, chief security officer at Corridor and former security lead at major tech companies, said large language models have reached or exceeded human capability for bug finding in many contexts.
But finding bugs is not the same as fixing them. Stenberg welcomes better detection tools but worries about maintainer capacity. Many critical projects are understaffed and underfunded, and a surge of findings can increase triage work. He believes AI is generally stronger at locating defects than at producing trustworthy fixes. Identifying a real problem and agreeing on a correct remedial approach often requires judgment, coordination, and context that go beyond an automated patch. Once a problem is validated, implementing the fix is usually straightforward; the time and effort are consumed by verification and consensus-building.
Not everyone agrees AI cannot help with repairs. Some vendors are developing agentic systems designed to autonomously find and mend vulnerabilities, indicating commercial interest in more end-to-end automation.
There is also a defense-versus-offense distinction to consider. Discovering a flaw is the first step in the kill chain; turning it into a working exploit is a separate, often technically demanding process. Leading commercial foundation models from top labs have safety guardrails and are proprietary, which makes them harder to repurpose for offensive uses. The larger concern is open-weight models, whose parameters are publicly available: if adversaries obtain or recreate high-capability open models and strip out safety measures, they could potentially both find bugs and generate exploit code. Some experts say the most advanced open-weight models trail the top closed models by less than a year, narrowing that gap.
Regulatory and policy debates are already surfacing. The U.S. Department of Defense designated Anthropic a supply-chain risk, potentially limiting government and contractor use of its technology; Anthropic is contesting that designation in court. Proponents of limited releases argue that controlled access gives defenders time to adapt and strengthen defenses while research continues.
In short, the rapid improvement of AI in security research is reshaping how vulnerabilities are found and handled. It promises faster and broader detection of flaws, but also raises new challenges: triaging higher volumes of reports, prioritizing limited maintainer time, ensuring fixes are correct, and preventing misuse as more capable models become accessible. How the industry balances accelerated discovery with responsible deployment and robust defenses will shape the security landscape in the years ahead.