← sources

Whatever happened to "Security Nightmare" Mythos?

source: reddit (reddit:ClaudeAI)published 2026-10-02 06:50ingested 2026-10-03 02:36

retained text

Back in April, Anthropic dropped a bombshell PR statement where they claimed their new model "Mythos preview" was too hot to handle because it was best of the best in hunting and exploiting vulnerabilities in code.

Being the responsible AI developers, they were not releasing it publicly but will be sharing its "power" with a select group of companies to review their (often closed) source codes and patch them before bad actors abused it. This was dubbed Project Glasswing.

They also disclosed multiple vulnerabilities, including a long standing FreeBSD networking bug.

Its been almost 6 months and Mythos has come (and Fable thereafter), but the predicted vulnerability explosion is nowhere to be seen. Many bugs that have since been disclosed were mostly in obscure or very less used options or features of software. A couple of LPE in Linux (copy fail et al) were AI assisted but not exclusively Mythos thing.

Has there been any worthwhile contribution of project glasswing (cynic: except that it has perhaps allowed Anthropic to train Fable on source codes that otherwise would not be available to it)?

Edit: A lot of people are commenting on the scale of vulnerabilities (specifically CVE entries) reported this year compared to prior years. The contribution of ALL AI models to that is not in question. Open weight and properietary AI models (inc Anthropic ones) have generated many of those CVEs. My question is specifically if Mythos has actually delivered (in terms of finding vulnerabilities) what was hyped in its release blog post?

I also now understand that Fable is Mythos but with guardrails around cyber security and bio research and that Mythos isn't generally available. I can see why Anthropic may be jumpy about bad actors exploiting its ability to chain exploits into usable cyber kill chain, but code analysis capability should be delivering top quality CVEs in large numbers with CVSS scores above 7 if it lives up to release hype. I think Browser sandboxes are one area that have benifitted most from AI review because of how their codebase is. But most of them are open source. Has Mythos proved to be better than other AI models in discovering high quality bugs there?

document fetched 2026-10-03 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1wvmwys/whatever_happened_to_security_nightmare_mythos