"The advantage will belong to the side that can get the most out of these tools," Anthropic warned — a blunt acknowledgement that its new research into frontier AI for code may reshape who controls offensive and defensive cyber capabilities.
Anthropic's Mythos model: a frontier for code reasoning and autonomy
On April 7 Anthropic announced an early preview of a model it calls Mythos, describing it as a "new frontier model" with "strikingly advanced capabilities in computer security tasks." According to the company, Mythos demonstrates major improvements in code reasoning and autonomy that place it well above the firm's then-flagship model, Opus 4.7. Anthropic said those improvements allow the model to automatically develop functional cyberattacks at a highly professional level — a capability the company characterized as a severe risk to private and public software.
claude-mythos-1-preview surfaced briefly in Claude Code and Claude Security
Evidence that Anthropic is preparing Mythos for a controlled public rollout appeared in product interfaces. The model identifier claude-mythos-1-preview showed up in Claude Code and briefly in Claude Security. Some users noticed a toggle to enable Mythos in the public version of Claude Code before the control was taken offline. That appearance confirms preparation for broader availability but Anthropic has not signaled whether Mythos will be available across all subscription tiers.
Glasswing: working with partners to mitigate AI-driven exploits
Anthropic said it is running a project named "Glasswing" to collaborate with other companies to secure high-value software against potential AI-driven exploits. Glasswing uses the unreleased Mythos Preview and, Anthropic reported, has already helped up to 50 organizational partners. The initiative appears aimed at applying Mythos's capabilities to defensive tasks while keeping its most dangerous outputs from public misuse.
10,000 high- or critical-severity vulnerabilities found in the first month
Anthropic disclosed that Mythos managed to uncover 10,000 high- or critical-severity vulnerabilities in its first month of use. That tally — and the company's warning that attackers could gain the short-term advantage if frontier labs are careless — helps explain why Anthropic held back on a public release until it built stronger guardrails. Anthropic specifically cited the risk that attackers could exploit a massive volume of unpatched vulnerabilities in popular applications, naming Firefox as an example.
What this means for defenders, attackers, and enterprise security teams
- Defenders and security teams: Anthropic predicts that in the long term "defenders" will more efficiently direct resources and use these models "to fix bugs before new code ever ships." The company's Glasswing collaboration and the integration tests inside Claude Code and Claude Security suggest a defensive-first release strategy tied to partner work and guardrail development.
- Attackers and opportunistic actors: Anthropic warned that "in the short term, this could be attackers, if frontier labs aren’t careful about how they release these models." The model's demonstrated ability to produce functional cyberattacks and its early discovery of tens of thousands of severe vulnerabilities create a window in which misuse could amplify threat actor capabilities.
- Enterprise procurement and platform customers: The brief, removed toggle and the unclear availability across subscription tiers leave customers uncertain about access. Organizations using Claude Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, or Haiku 5.5 — models Anthropic currently offers — will be watching whether Mythos becomes a guarded add-on, a partner-only tool under Glasswing, or a more widely distributed product.
Anthropic's public language frames Mythos as both an unprecedented capability and a responsibility problem: its discoveries of vulnerabilities are already significant enough to delay release while guardrails and partner programs are developed. The brief, public-facing traces of claude-mythos-1-preview in Claude Code and Claude Security underscore that the company is moving from experimentation toward controlled deployment, yet the company has left open which customers will ultimately see the model.
For now, Anthropic's next steps — maturing guardrails, expanding Glasswing partnerships beyond the 50 partners it has already assisted, and deciding the model's availability across subscription tiers — will determine whether Mythos becomes primarily a defensive force multiplier or a short-term accelerant for attackers. The one concrete metric the company supplied — 10,000 high- or critical-severity vulnerabilities discovered in a month — makes clear why Anthropic chose caution.
