Anthropic’s Claude Mythos Generates Working Zero‑Day Exploits

Anthropic has released a preview of its new large language model, Claude Mythos, which can identify and exploit zero-day vulnerabilities in major operating systems when directed by users¹. In testing it obtained local privilege escalation exploits on Linux and produced a fully functional remote code execution payload targeting FreeBSD’s NFS service that granted unauthenticated users root access¹. Non-experts have found complete working exploits overnight after asking the model for them, indicating that the system can translate high-level vulnerability descriptions into executable attack vectors¹.

In response to these capabilities, Anthropic launched Project Glasswing aimed at securing critical software using this technology. The initiative plans to automate detection, patching or mitigation strategies derived from the same chain-of-thought reasoning that powers the exploit generation process¹.

The underlying mechanism—systematic application of chain‑of‑thought reasoning against known software patterns—is not entirely new. Back in 2018 at DefCon's Pwn Village researchers demonstrated automated fuzzing using LLMs for finding bugs; this work showed that a model could learn from feedback and iteratively improve its exploit generation within hours compared with manual methods taking weeks or months.

Zero‑day vulnerabilities are traditionally discovered through painstaking reverse engineering and manual code review; now an AI can perform this work autonomously, potentially exposing many more systems per week than human teams alone could uncover. By lowering technical barriers for non‑experts, the model enables threat actors lacking deep expertise to launch sophisticated attacks with minimal effort, amplifying the scale of potential compromise across industries such as healthcare, finance, and infrastructure.

Anthropic’s statements emphasize responsible use, noting that Project Glasswing will focus on patch automation rather than merely detecting flaws¹. Security practitioners warn that any system capable of generating exploit code must be tightly controlled, as misuse or accidental release could accelerate widespread exploitation¹.

Organizations should reassess their vulnerability management workflows to incorporate AI-driven scanning tools that mirror Mythos’ reasoning patterns while maintaining strict access controls and audit trails¹. The emergence of models like Claude Mythos requires defenders to adopt proactive mitigation strategies at unprecedented speed