Anthropic’s Mythos AI Is Too Dangerous to Release — So It’s Keeping It Secret

Anthropic’s Mythos AI model represents an unprecedented step in artificial intelligence capability — and, for the first time, a major AI lab is withholding its most powerful model from the public entirely. The decision, driven by alarming cybersecurity risk assessments, marks a turning point in how the industry handles models that cross new capability thresholds. Rather than a phased rollout, Anthropic launched Project Glasswing to grant tightly controlled access to select defensive-use partners while committing $100M in compute credits to ensure the technology is used responsibly.

What Is Anthropic’s Mythos AI Model?

Mythos Preview is Anthropic’s latest frontier model, internally classified as its most capable system to date. Unlike Claude 3.7 Sonnet or previous public releases, Mythos was developed with a sharper focus on reasoning across complex, multi-step problems — including those with real-world operational implications. Early internal evaluations showed the model could execute tasks in cybersecurity domains that previous models could not.

The model’s System Card — Anthropic’s transparency document outlining a model’s capabilities and risk profile — flagged something the company had never encountered before: credible potential for full automation of offensive cyber operations. That assessment triggered an internal review that ultimately led to the decision to withhold the model from standard deployment channels.

Why Mythos Is Too Dangerous to Release

The core concern isn’t that Mythos can answer cybersecurity questions — many existing models can do that. The issue is the degree to which Mythos can plan, sequence, and execute offensive operations with minimal human oversight. According to the System Card findings cited by NBC News experts, the model demonstrated the ability to identify vulnerabilities, generate working exploit code, and iterate on attack chains in ways that could accelerate real-world cyberattacks by months.

This isn’t theoretical. Security researchers who reviewed the System Card noted that existing red-team defenses were developed before models of this capability class existed. Releasing Mythos publicly — even with standard API guardrails — would, in Anthropic’s assessment, provide a meaningful uplift to malicious actors that current defensive infrastructure isn’t equipped to absorb.

Project Glasswing: Restricted Partner Access

Project Glasswing is Anthropic’s structured response to the Mythos dilemma: rather than shelving the model entirely, the company is routing access exclusively through vetted defensive cybersecurity organizations. The initial partner cohort includes CrowdStrike and Microsoft, both of which operate large-scale threat intelligence and incident response operations where Mythos’s capabilities could be applied to detect and neutralize attacks rather than enable them.

Access under Glasswing is not a standard API key. Partners undergo vetting, agree to usage constraints, and operate within monitoring frameworks that allow Anthropic to audit how the model is being used. The program is explicitly defense-only — partners cannot use Mythos for offensive testing or red-team exercises without explicit case-by-case approval.

The $100M Compute Fund and Open-Source Grants

Alongside the Glasswing partner program, Anthropic announced a $100M usage credits fund for qualifying cybersecurity organizations. The credits allow eligible defenders — including non-profits, government agencies, and academic security research teams — to access Anthropic’s broader API suite for threat detection, vulnerability research, and defensive tooling development. This is separate from Mythos access, which remains under stricter Glasswing controls.

Complementing the compute fund, Anthropic committed $4M in open-source security grants to support community-led projects that build detection and defense capabilities using publicly available models. The goal is to raise the floor of defensive capability across the ecosystem, even for organizations that will never have Glasswing-level access to Mythos itself.

Frequently Asked Questions