What is Claude Mythos: An AI Cybersecurity Researcher

Claude Mythos is the AI that found thousands of zero-day flaws. This AI model is either the most breakthrough in cybersecurity history, or Anthropic’s marketing team has crafted the most convincing narrative of the decade. The truth lies somewhere in between, and the stakes are far higher than most people realize.

On June 9, 2026, Anthropic released Claude Fable 5 to the public, a model in the same family as Claude Mythos 5, which the company had previously declared “too dangerous to release” altogether. Claude Mythos Preview first appeared in April 2026, claiming it could autonomously discover and exploit zero-day vulnerabilities across all major operating systems and browsers.

Claude Fable 5: Publicly Accessible

Claude Fable 5 represents Anthropic’s first publicly available Mythos-class model, released simultaneously with the restricted Claude Mythos 5 on June 9, 2026. Both models share the same underlying architecture as Claude Mythos Preview, announced on April 7, 2026.

Claude Fable 5 includes guards that block responses in high-risk areas like cybersecurity and biology, making it suitable for general public use. In contrast, Claude Mythos 5 is a security-removed version released exclusively to vetted organizations, including U.S. government cyber defenders, through Project Glasswing.

Anthropic justified the broad release of Fable 5 through new safeguards that prevent responses in specific high-risk areas. The company will retain conversations between users and AI models for 30 days, with limited reviewers able to access conversations “if a potential for significant harm is identified or if there is a written request from the customer”.

Microsoft has already restricted Claude Fable 5 for internal employee use pending data retention policy evaluations, with the legal team assessing whether Anthropic’s requirements align with Microsoft’s internal standards. Microsoft made Claude Fable 5 available to GitHub Copilot and Foundry customers but excluded it from the internal model selection screen used by employees.

Claude Mythos 5: Restriction

Claude Mythos 5 operates at a capability tier above Opus, representing the first public version of this Mythos-class model with some cyber-related safety measures removed. The model is positioned as an updated version of Claude Mythos Preview, with restricted access limited to vetted partners.

The pricing structure for Claude Mythos 5 reflects $10 per million tokens for input and $50 per million tokens for output, with a 1-million-token context window. Access remains limited to organizations that pass Anthropic’s vetting process, maintaining the company’s original stance that the technology requires controlled deployment.

ExploitBench testing shows Claude Mythos 5 achieving 78% exploit success rates for cyber defense applications. This benchmark measures the model’s ability to identify and exploit vulnerabilities, providing concrete metrics for its cybersecurity capabilities.

What is Project Glasswing

Project Glasswing represents Anthropic’s strategy for managing Mythos’s dual-use nature, the same technology that can defend systems can also attack them. The initiative involves approximately 40 to 200 high-profile organizations receiving limited access to the Claude Mythos Preview for defensive cybersecurity work.

Major partners include AWS, Apple, Microsoft, Google, and Nvidia, representing the technology sector’s most influential companies. Fortune first reported that Anthropic was developing and testing the model, describing it as “by far the most powerful AI model” the company had ever developed

Anthropic won’t release Mythos publicly but is working with selected firms to explore how it could prevent cyberattacks. This approach allows defensive security teams to leverage Mythos’s capabilities while preventing unrestricted access that could enable malicious actors.

Bain & Company describes Claude Mythos and Project Glasswing as “a business risk of the highest order, not a technology problem to be delegated downward”. This framing emphasizes that organizations must treat AI cybersecurity capabilities as strategic business risks requiring executive-level attention.

Zero-Day Flaws: Case Study

Claude Mythos Preview’s most important achievement is its ability to autonomously discover zero-day vulnerabilities that survived decades of expert review. The AI identified critical flaws, including a 27-year-old OpenBSD vulnerability and a 16-year-old FFmpeg flaw, plus Linux kernel vulnerabilities.

On the Firefox 147 JavaScript engine benchmark, Mythos developed working exploits 181 times compared to just 2 for Claude Opus 4.6, representing a 90x improvement. Opus 4.6 turned Firefox 147 vulnerabilities into shell exploits only two times out of several hundred attempts, while Mythos Preview achieved register control on 29 additional attempts beyond the 181 successful exploits.

Mozilla released Firefox 150 with 271 fixes credited to Anthropic’s Claude Mythos Preview, though MFSA 2026-30 publicly attributes only three CVEs to researchers using the model. This discrepancy suggests Mythos identified far more vulnerabilities than were officially documented in Mozilla’s release notes.

The red team blog walks through specific findings in detail, including a 27-year-old OpenBSD TCP SACK bug involving signed integer overflow, a 16-year-old FFmpeg H.264 vulnerability, a guest-to-host memory corruption bug in a production memory-safe virtual machine manager, and fully autonomous exploitation of a 17-year-old FreeBSD NFS remote code execution vulnerability (CVE-2026-4747) including a multi-packet ROP chain.

Mythos found vulnerabilities in every major operating system and every major web browser, including chaining four vulnerabilities into a JIT heap spray that escaped both renderer and OS sandboxes. These capabilities were not explicitly trained; they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.

The model saturated Cybench with a 100% solve rate on all 35 Capture-The-Flag challenges, rendering the benchmark obsolete. Additional benchmark numbers include 93.9% SWE-bench Verified, 77.8% SWE-bench Pro, 82% Terminal-Bench 2.0, and 97.6% USAMO 2026.

Is This Real or Marketing Hype

Despite Anthropic’s impressive claims, credible voices in the security community question whether Mythos represents the breakthrough the company describes. Bruce Schneier, renowned security cryptographer and privacy activist, explicitly stated that Anthropic “might not be” as amazing as claimed.

The UK Government’s AI Security Institute conducted independent testing and found that Mythos “was not significantly better at individual cybersecurity tasks compared to other models”. This finding directly contradicts Anthropic’s positioning of Mythos as uniquely powerful for cybersecurity applications.

BBC News reported that “the company claims the AI tool can outperform humans at some hacking and cyber-security tasks,” but emphasized the need to understand what risks it poses. The article frames Mythos as controversial rather than universally accepted as a breakthrough.

Security researcher John Z. Black’s full analysis emphasizes that while Mythos demonstrates unprecedented vulnerability discovery, the technology’s real-world impact requires careful evaluation. The analysis doesn’t dismiss Mythos’s capabilities but calls for measured assessment rather than hype-driven adoption.

What This Means for Cybersecurity Professionals

The emergence of Claude Mythos forces a fundamental reckoning in cybersecurity. If an AI can autonomously discover vulnerabilities that survived 27 years of expert review, the traditional patch-and-defend model becomes unsustainable. Defenders face an asymmetric flood of bugs that could drown security teams.

For organizations, Claude Mythos represents both a defensive tool and an existential threat. The same model that Project Glasswing partners use to identify vulnerabilities in their systems can be wielded by attackers to exploit those same vulnerabilities. 

Will It Effect Bug Bounty Hunting

The emergence of Claude Mythos sends waves through the bug bounty industry, but the reality is more nuanced than the “AI will destroy bug bounty hunting” narrative suggests.

Mythos is a sign that the buffer is collapsing. As discovery and exploitation converge, vulnerability workflow becomes a speed problem.

AI Makes Bug Hunting Cheap

The most immediate impact is economic: LLM bug hunting is pretty cheap, meaning 1-person projects can do security testing themselves without contracting out to expensive security firms.

Some experts argue that Claude Mythos will destroy the bug bounty industry, arguing that these critics fail to understand why platforms really work. A platform’s strength comes from more than one hacker’s skills; it relies on diverse perspectives, specialized expertise, and the human ability to understand business logic that AI may miss.

AI’s impact on bug bounty programs includes lowering the cost to finding bugs, making people more likely to participate in bug bounty programs or even report them for free. This could expand the pool of hunters, particularly among smaller developers and security enthusiasts who previously lacked the resources for deep security research.

YesWeHack has launched and optimized dozens of Bug Bounty Programs with AI-related scopes, from simple chatbot interfaces to complex multi-system architectures, showing that AI scopes themselves are becoming part of bug bounty programs.

Bottom Line

What remains clear is that the world now has access to AI that can find vulnerabilities in Firefox 181 times more effectively than previous models, identify 27-year-old bugs that survived decades of review, and chain multiple vulnerabilities into exploits that escape sandboxes. Whether this technology strengthens or weakens global cybersecurity depends entirely on how organizations, governments, and individuals choose to deploy it.

FAQ