Mozilla released Firefox 150 on April 21, 2026, claiming Anthropic’s Mythos Preview identified 271 vulnerabilities, yet the official security advisory lists only 3 CVEs credited to the AI tool. This dramatic discrepancy—a factor of 90—has sparked confusion in the cybersecurity community about how AI-discovered vulnerabilities are counted and disclosed.
According to Bobby Holley’s Mozilla blog post, Firefox 150 ships with fixes for 271 vulnerabilities found by Claude Mythos Preview. However, the canonical security advisory MFSA 2026-30 tells a different story, listing only 41 total CVEs with just three carrying the “using Claude from Anthropic” credit.
The Numbers Don’t Add Up
The discrepancy becomes more puzzling when compared to previous releases. Firefox 148’s February release showed perfect alignment between Mozilla’s blog claims and official advisories. MFSA 2026-13 listed 51 CVEs total, with 22 credited to Anthropic—exactly matching the blog post’s count.
Anthropic’s own red team analysis of the Opus 4.6 round revealed important context:
- 112 bug reports submitted to Firefox
- 100% confirmed as true positives
- 22 shipped CVEs in the final release
This created a roughly 5:1 submission-to-CVE ratio, explained by duplicate findings, issues fixed without CVE assignment, and vulnerabilities rolled into omnibus memory-safety entries.
Mythos Preview’s Limited Public Footprint
Despite claims of 271 vulnerabilities, analysis by flyingpenguin found minimal evidence in Firefox 150’s public record. The security advisory shows:
- 41 total CVE entries
- 3 standard memory-safety roll-ups
- Only 3 individual CVEs credited to Claude from Anthropic
If the same 5:1 ratio applied to Firefox 150, 271 submissions should have produced approximately 54 CVEs, not 3. The actual ratio jumps to an unprecedented 90:1, raising questions about the nature of Mythos Preview’s findings.
AI Vulnerability Hunting Capabilities
Mozilla’s experience highlights the potential impact of advanced AI tools on cybersecurity. According to Wired, Firefox CTO Bobby Holley believes these tools represent a dramatic shift in vulnerability discovery capabilities.
“Our belief is that the tools have changed things dramatically, because now we have automated techniques that can cover, as far as we can tell, the full space of vulnerability-inducing bugs,” Holley stated.
The Firefox team acknowledges the challenge of managing the “firehose of bugs” that AI tools can generate, requiring significant resources and process adjustments. However, they view this investment as necessary given that similar capabilities will inevitably reach attackers.
Limited Access and Government Use
Anthropic has restricted Mythos Preview access to approximately 40 organizations due to the model’s offensive cybersecurity capabilities. TechCrunch reports that the National Security Agency is among the undisclosed recipients, using Mythos primarily for vulnerability scanning despite ongoing Pentagon disputes with Anthropic.
The UK’s AI Security Institute has also confirmed access to Mythos Preview, highlighting international interest in the technology. This limited distribution reflects growing concerns about AI models’ dual-use potential in cybersecurity.
Industry Implications and Concerns
The Mozilla case illustrates broader challenges facing the cybersecurity industry as AI capabilities advance. Both Anthropic and OpenAI have announced models with enhanced cybersecurity features, prompting industry working groups to assess implications and develop strategies.
Key considerations include:
- Transparency in vulnerability disclosure when AI tools are involved
- Standardization of counting methodologies for AI-discovered bugs
- Resource allocation for processing increased vulnerability volumes
- Timeline pressures for patching AI-discovered flaws before attackers gain similar capabilities
What This Means
The Mozilla Mythos discrepancy reveals critical gaps in how the industry handles AI-assisted vulnerability discovery. The 268-vulnerability difference between claimed findings and official CVEs suggests either fundamental counting methodology changes or transparency issues in disclosure processes.
This case establishes important precedent as more organizations adopt AI vulnerability hunting tools. Clear standards for reporting AI-discovered vulnerabilities will become essential for maintaining trust and enabling proper risk assessment across the cybersecurity ecosystem.
The dramatic efficiency difference between Firefox 148 and 150 releases also raises questions about Mythos Preview’s actual capabilities versus marketing claims, potentially influencing future AI security tool evaluations.
FAQ
Q: Why is there such a large discrepancy between claimed and documented vulnerabilities?
A: The 271 claimed vulnerabilities likely include duplicates, non-CVE issues, and findings rolled into broader security fixes, while only 3 received individual CVE assignments with Anthropic credit.
Q: How does this compare to previous AI vulnerability discoveries?
A: Firefox 148 showed perfect alignment with a 5:1 submission-to-CVE ratio, making Firefox 150’s 90:1 ratio highly unusual and potentially indicating methodology changes.
Q: What are the implications for cybersecurity transparency?
A: This case highlights the need for standardized reporting of AI-discovered vulnerabilities to maintain public trust and enable proper risk assessment as these tools become more widespread.
Sources
- Mythos Mystery in Mozilla Numbers: How 22 Vulns Became 271 or Maybe 3 in April | flyingpenguin – www.flyingpenguin.com
- Mozilla Used Anthropic’s Mythos to Find and Fix 271 Bugs in Firefox – Wired
- Mythos remains a mystery as security world faces rising threats, agentic attacks and concerns about AI integrity – SiliconANGLE – Google News – AI Security
- NSA spies are reportedly using Anthropic’s Mythos, despite Pentagon feud – TechCrunch
