Anthropic Mythos Preview AI Reveals Critical Cyber Flaws

Anthropic’s Claude Mythos Preview is a general‑purpose AI capable of identifying thousands of previously unknown software vulnerabilities and generating proof‑of‑concept exploits. Access is tightly limited under Project Glasswing to reduce the high risk of malicious misuse.

Anthropic reports that Mythos goes beyond earlier models in code reasoning and autonomous security work: it recognizes weaknesses, reproduces them and can create working proof‑of‑concept exploits. In Anthropic’s internal tests the model achieved an 83.1% success rate on first attempts for those proof‑of‑concepts, and found flaws in major projects including the Linux kernel and older OpenBSD issues (Anthropic; Axios).

News coverage and company statements portray Mythos as far more effective than predecessors. For example, reporting contrasts Mythos with a prior model, Opus 4.6, which reportedly found about 500 zero‑days in open‑source software; Mythos’ results were described as much larger in scale (Axios).

Not all circulating figures are verified. Claims such as “one team in seven weeks found more than 2,000 zero‑day vulnerabilities” or comparisons equating Mythos’ haul to a large share of global annual zero‑day output were not substantiated in Anthropic’s public materials or the reporting reviewed. Independent sources did not confirm specific numbers or certain attributions. The verifiable facts remain: Mythos found many previously unknown vulnerabilities and Anthropic is keeping details private to limit abuse (Anthropic; Axios; YouTube report).

Anthropic restricted Mythos access to a select set of industry and security partners — roughly 40 companies and organizations — spanning big tech, security vendors, financial institutions, chip makers and open‑source foundations. The program, called Project Glasswing, reportedly includes up to $100 million in usage credits to encourage defensive testing under managed conditions (Axios; Anthropic).

Anthropic and others stress that a broadly available model capable of generating exploit code would let bad actors automate discovery and exploitation at scale. That dual‑use threat is the primary reason Anthropic opted for controlled access rather than a public release (Anthropic).

A YouTube report alleges that unauthorized users obtained access to Mythos shortly after Anthropic’s announcement. The video claims the leak exploited perimeter and operational weaknesses rather than a flaw in the model itself. Reportedly the unauthorized users did not launch known attacks, but the episode highlights that guarding access to a vulnerability‑finding tool is itself a security challenge.

Security leaders describe this as a potential turning point: AI can now find and exploit flaws more quickly and at a larger scale than many human teams. That raises urgent questions about defense assumptions and where security budgets should go (Anthropic; Axios).

Experts recommend a strategic shift away from relying primarily on perimeter tools toward assuming breaches are possible and protecting data directly: stronger encryption, stricter key management, least‑privilege access and data‑level safeguards so that stolen access does not automatically yield readable data.

Anthropic is not alone. Reporting indicates rivals, including OpenAI, are developing comparable restricted models and programs (for example, OpenAI’s reported “Trusted Access for Cyber”), and industry watchers expect similar capabilities to be introduced within months. That could accelerate the pace at which new vulnerabilities are discovered and exploited (Axios).

Anthropic used coordinated disclosure to notify affected maintainers and intentionally withheld technical details because more than 99% of the vulnerabilities it found remain unpatched. The company frames Mythos as a defensive tool but acknowledges the inherent dual‑use tension: public safety is served when maintainers learn about and patch serious bugs, yet broad release of an exploit‑generating tool risks enabling widespread attacks (Anthropic).

Economic and infrastructure risk: Rapid AI‑driven discovery of deep software flaws could expose weak points across banking, utilities, transportation and healthcare systems that rely on common software stacks, creating short windows of heightened global risk (Anthropic; Axios).

Policy and regulation: Governments may face pressure to define rules for responsible access to vulnerability‑finding AI. Private programs like Project Glasswing are one approach, but public policy and international coordination will likely be required to set oversight, disclosure obligations and cross‑border cooperation (Anthropic).

Security strategy shift: Organizations should reassess spending priorities toward data‑level protection: encryption, key management, least‑privilege and designs that limit post‑breach impact. Cyber insurers, auditors and boards must incorporate AI‑driven risk into policy and budgets.

Research and innovation: Safely adopting models like Mythos will require improved guardrails, auditing tools and secure operational practices. Controlled access programs can help, but leaks illustrate the need for tight operational controls and independent oversight (Anthropic; YouTube report).

Practical advice: Security teams should inventory critical assets, harden encryption and key controls, review privileged access, accelerate patching, and maintain strong monitoring for unusual, automated exploit activity. Partnerships among vendors, maintainers and trusted AI programs will be critical to coordinate disclosure and mitigation.

This article is based on Anthropic’s Mythos Preview report and media coverage. Key sources include Anthropic’s technical post, Axios’ reporting, and a YouTube report on alleged unauthorized access. Some online claims about exact zero‑day counts and attributions were not corroborated and remain unverified.