Claude Mythos: A New Frontier in AI Cybersecurity
Anthropic's Claude Mythos is finding vulnerabilities at a scale no human team can match. Whether that speed helps defenders more than attackers is the question nobody can answer yet.
Key points
- Anthropic released Mythos 5 in April 2026 to a vetted group of technology partners for cybersecurity and healthcare applications.
- Project Glasswing's 50 initial partners used it to find over 10,000 high- or critical-severity flaws across major operating systems and browsers.
- Export controls on Mythos 5 and Fable 5 took effect 15 June and were lifted 15 days later on 30 June.
- Claude Fable 5 offers a guardrailed alternative, though critics say those guardrails trip on routine tasks.
Anthropologic has been busy. Anthropic, the company behind the Claude family of models, launched Mythos 5 to a small pool of vetted partners before a wider rollout. The access scheme, called Project Glasswing, pulls in infrastructure providers, open-source developers, and large technology companies. The goal: let defenders find and patch flaws before attackers, who are also leaning hard on AI, can exploit them.
The early numbers are hard to ignore. Those 50 initial partners turned up more than 10,000 high- or critical-severity flaws spanning every major operating system and browser, including a 27-year-old bug in OpenBSD, the security-focused Unix variant. Mythos can also chain separate flaws together into a single attack sequence, which is where frontier AI starts to look genuinely different from prior tooling.
In June, 150 additional organizations were granted access. All partners must accept a 30-day data-retention policy for safety monitoring. When Mythos availability pushed the Trump administration to revisit its light-touch AI oversight stance, the US government applied export controls to both Fable 5 and Mythos 5 on 15 June. Those controls were lifted on 30 June.
For broader audiences, Anthropic offers Fable 5, built on the same underlying technology but with guardrails that block operations in high-risk security domains. Flagged queries are routed instead to the older Opus 4.8 model. The guardrails have frustrated practitioners: complaints circulating on security podcasts describe Fable 5 downgrading to Opus 4.8 when a user asks it to summarise a blog post or spell the word "exploit."
Cisco, a Project Glasswing partner, open-sourced its Foundry Security Spec, a model-agnostic test harness, so other teams can build comparable workflows without starting over. Cisco argues internally that AI lets teams identify and resolve vulnerabilities far faster than the old find-one-patch-one model allows, and it's scanning 1.8 billion lines of its own code using the technology. Smaller organisations, Cisco notes, don't need restricted models: tightening authentication, segmentation, and remediation of actively exploited vulnerabilities will move the needle more.
Should you worry?
Five Eyes intelligence agencies warned in a joint statement that frontier models like Mythos are reshaping offensive and defensive capabilities "in a scale of months rather than years." Joe Hubback, partner and CISO at consultancy Elixirr, put it plainly: AI systems can now map attack paths across software and critical infrastructure faster than human adversaries can catalogue them, and broad commercial release of Mythos-class tools makes that every organisation's problem, not a niche research concern.
The honest critic's view, which we've tracked across three earlier Mythos stories since our first coverage on 11 June, is that volume vulnerability discovery solves only part of the problem. False positives, unreliable fixes, and social engineering sit entirely outside what any of these models address. Finding bugs faster is real progress. Knowing what to do with the list is the harder job.
