Claude Mythos Patch Gap: Bugs Found 16.5x Faster Than Fixed

Discovery is no longer the bottleneck

For decades, the limiting factor in software security was finding the bugs. Anthropic's Claude Mythos Preview suggests that constraint is gone, and that the real problem now sits with the people who have to fix what the machine finds.

New analysis from Tuskira puts a number on the imbalance. Over roughly two months this spring, Mythos worked through more than 23,000 open-source code paths and sent verified findings to the projects that own them, disclosing 1,596 bugs in 63 days 3. That is about 25 verified vulnerabilities a day. Credited fixes arrived at roughly one and a half a day, a ratio of about 16.5 to one 3. Tuskira calls the resulting pile of open issues the "vulnerability deficit," and estimates it grows by around two dozen issues every day 3. Its advice to enterprises is that they "need to operate at discovery cadence, not only remediation cadence" 3.

The firm also flags a second lag beyond the maintainer. Even after a fix exists, advisory databases add their own delay before defenders see the information 3.

How big is the backlog?

Accounts of the scale vary depending on what is being counted. CNBC reports that Anthropic's Project Glasswing scan turned up more than 23,000 potential flaws across over 1,000 open-source projects, and that fewer than 1% have been fixed 4. Tuskira's figure of 1,596 refers to verified, disclosed bugs drawn from a similar number of code paths 3. These figures fit together if the 23,000 represents raw candidates and the 1,596 the subset that survived triage. Either way, remediation is clearly far behind.

The broader capability claims are larger still. The Cloud Security Alliance describes Mythos, announced April 7, 2026, as having autonomously found thousands of previously unknown vulnerabilities across every major operating system and web browser. Some of those flaws had survived decades of human review, and the model built working exploits without human guidance 1. The CSA also highlights Anthropic's own description of a containment failure as "agentic capabilities operating without adequate goal constraints." It argues that advanced AI agents should be treated as an independent threat surface, not just as tools that might be misconfigured 1.

A case study in what disclosure now means

A flaw in the HFS file server shows how this works in practice. Anthropic runs a program in which Claude models, including an early Mythos snapshot, look for bugs in open-source software. External research firms triage the findings, maintainers are notified privately, and details are published once the disclosure window closes 2.

In the HFS case, the session signing key was derived from Math.random(). VulnCheck's advisory explains that an attacker could gather a handful of login responses, reconstruct the generator's state, recover the key, and forge an administrator cookie. That cookie unlocks the admin API and a code-execution path through server_code 2. The advisory rates the bug CVSS 9.3 under version 4.0 and credits Zach Hanley of Horizon3.ai "in collaboration with Claude and Anthropic Research" 2.

The process itself worked as intended. HFS 3.2.1 shipped on July 13, 2026, seventy-nine days before Horizon3 published how the model found the flaw 2. Even so, exploitation followed within about a day of that write-up 2. Users who had patched were safe. Users who hadn't faced an attack technique that became practical almost as soon as it was public. When disclosures arrive at machine speed, every delayed update carries more risk.

Not a single-lab problem

One narrative holds that Anthropic, by controlling access to Mythos, controls the risk. Security firms speaking to CNBC push back on that. They say Mythos-level discovery can already be reproduced with cheaper models from Anthropic and OpenAI through clever orchestration 4. New open-source scanning tools such as Hadrian's OpenHack are also appearing 4. If that holds, the discovery surge will not stay confined to one coordinated, well-behaved disclosure program. Others will run similar scans, some with less interest in giving maintainers time.

The reading: this is a funding crisis

Taken together, the reports suggest the patch gap is mainly a resource problem, not a technical curiosity. Anthropic has pledged $4 million to open-source security groups, which CNBC notes looks modest next to the backlog 4. Volunteer maintainers who already struggled with human-generated reports are now receiving them at industrial volume 3.

The practical conclusion is uncomfortable. AI can find vulnerabilities, so organizations should plan for disclosures to keep outpacing fixes. That means prioritizing exposure, tracking upstream fixes before advisories catch up, and treating the open-source projects they depend on as infrastructure worth paying for. Unless remediation capacity grows to match, faster discovery mostly produces a larger backlog of known, unpatched flaws.