Mythos Meets cURL: One Low-Severity Win, Plenty of Skepticism
When Anthropic promoted its Mythos AI as too potent for public release, expectations in the security world soared. cURL creator Daniel Stenberg was among those invited to try it—indirectly—through Anthropic’s Project Glasswing. Instead of hands-on access, he received a single scan report run by someone else against cURL’s master branch. The result: Mythos flagged five issues it labeled “confirmed security vulnerabilities.” After several hours of review by the cURL security team, four of those were downgraded. Three were false positives that restated limitations already documented in the API, and one was treated as a simple bug rather than a security flaw. Only one issue survived as a genuine vulnerability, and even that will be disclosed as a low-severity CVE in an upcoming release. For Stenberg, the outcome underscored a blunt conclusion: Mythos may do competent AI security testing, but its performance looks comparable to existing tools rather than revolutionary.
Mozilla’s Firefox Experiment: AI, Middleware, and 423 Fixed Bugs
Mozilla’s experience with Anthropic Mythos tells a more optimistic story—at least at first glance. In April, Firefox engineers report fixing 423 security bugs, compared with just 76 in March and a long-term monthly average of 21.5. They credit AI security bug detection with a big share of that surge, saying Mythos Preview alone identified 271 vulnerabilities in Firefox 150, backed by Anthropic’s Opus 4.6 model. However, Mozilla’s team is careful to highlight that the real breakthrough may be the “agentic harness,” the middleware layer that orchestrates how the models scan code, reason about findings, and filter noise. Recent AI-generated reports have shifted from messy to genuinely useful as this harness improved. The sample bugs Mozilla publicly linked include a 20-year-old high-severity heap use-after-free and several sandbox escapes that typical fuzzing misses. Still, outside observers note that it’s difficult to separate Mythos’s raw capability from better tooling and workflows wrapped around it.
Bypassing macOS Security: Mythos and the Chained-Exploit Demo
Beyond open source projects, Mythos has also been credited with helping security researchers break into macOS in a novel way. A Palo Alto-based research team reportedly used the Anthropic Mythos model to design a chained attack that linked two macOS bugs to corrupt system memory. Once memory was compromised, the exploit could access parts of the device that should have remained off-limits, raising the possibility of gaining deeper control when combined with other techniques. The findings impressed the researchers enough that they traveled to Apple’s headquarters to brief the company directly. Apple says it is reviewing and validating the reported issues, though it has not publicly detailed or confirmed any fixes. Importantly, the researchers themselves emphasize that Mythos did not operate autonomously. The exploit required human expertise to guide, refine, and safely execute the AI-derived strategy, highlighting that Mythos acts more like a powerful assistant than a fully automated hacking engine.
Model vs. Marketing: What Mythos Really Changes in Security Testing
Across cURL, Firefox, and macOS, a pattern emerges: Anthropic’s Mythos model is capable, but its impact is uneven and context-dependent. In highly scrutinized code like cURL, Mythos produced one low-severity vulnerability and several non-security bugs—useful, but far from the sweeping breakthroughs suggested by Anthropic’s rhetoric. In Firefox, the spike in fixes appears tied not only to Mythos itself, but to Mozilla’s careful integration, triage, and custom middleware that boosts signal over noise in software vulnerability scanning. Meanwhile, the macOS work shows Mythos can help craft sophisticated chained exploits, yet only in tandem with seasoned human researchers. That combined evidence fuels skepticism in the security community about whether marketing has outpaced reality. Mythos clearly advances AI security testing, but for now it looks less like a singular game-changer and more like another strong model whose real power depends on how teams integrate, supervise, and verify its results.