Anthropic’s CEO Just Issued a 6-Month Warning

Anthropic CEO Dario Amodei walked on stage this week and dropped a number that should make every CIO uncomfortable: tens of thousands of unpatched software vulnerabilities, surfaced by an internal Anthropic model that is officially still in testing. The model is called Mythos, and according to Amodei, it has compressed a decade of vulnerability research into a few months of automated scanning.

A Number Nobody Wants on the Front Page

From 20 Firefox bugs to 300 — in one model upgrade

Anthropic’s previous internal models had already surprised security teams by finding around 20 vulnerabilities in Firefox during scoped testing. Mythos, the company’s most capable cybersecurity-tuned model to date, jumped that figure to roughly 300 issues in the same browser. Extrapolated across the rest of the open and proprietary software stack, the Mythos sweep has produced what Amodei describes as tens of thousands of distinct findings, the vast majority still unpatched and undisclosed.

The reason for the silence is straightforward. Disclosure rules and basic risk management both say: do not publish a working exploit recipe before the fix is shipped. So the public count is small. The internal count, the one Mythos is actually generating, is what Amodei is asking the industry to reckon with.

Project Glasswing: Who Gets the Keys

A short, deliberate guest list

Mythos is not a public product. Under an internal program called Project Glasswing, Anthropic has granted vetted access to a narrow group of organizations whose job is to fix the things Mythos is finding. Reported participants include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft and Nvidia. The list reads like a roll call of the platforms most of the global software supply chain depends on.

The structure mirrors how nuclear physics labs handle dual-use research: capability is shared, but only with parties who have the capacity to mitigate the downside before the upside becomes a weapon. For Anthropic, this is also a hedge against a known leak risk — earlier this year, a small group of users reportedly guessed where an internal Mythos checkpoint was hosted and accessed it before the company shut the door.

The 6-Month Window

“A moment of danger” — and a clock

Amodei’s framing is not “we are doomed.” He calls the current period a moment of danger, and explicitly says it ends well if the response is fast. His estimate: institutions have roughly six to twelve months to patch the bulk of what Mythos is uncovering before models of comparable capability reach actors who do not share the same disclosure norms. The reference to Chinese AI labs is not subtle. The implicit assumption is that frontier capability gaps in cybersecurity-relevant tasks compress faster than gaps in, say, multimodal reasoning.

Sitting next to JPMorgan CEO Jamie Dimon at a recent event, Amodei was asked whether the cyber freakout was warranted. Both men sidestepped the binary question and instead pivoted to what enterprises should be doing right now: inventorying their software bill of materials, prioritizing patch cadence on browser engines, identity providers, and core OS components, and treating any unsupported library as already compromised.

What This Means for Builders

Three concrete shifts to expect

First, expect a wave of silent CVEs. Browser, kernel and runtime vendors will publish fixes faster than usual, with terse advisories that refer back to internal AI-assisted reviews. Second, expect insurance and procurement contracts to start asking a new question: which AI security tooling do you have access to, and how often does it run against your stack? Third, expect open-source maintainers to receive a flood of pre-disclosed reports they cannot triage alone, accelerating the push for AI-assisted maintainer tooling funded by the same Glasswing-tier companies.

For startups building in the AI-for-security space, the message is mixed. The market just got bigger and more urgent, but the largest deal flow is concentrating around the very small set of vendors who can credibly say their tooling rivals what Mythos is doing inside Anthropic. Differentiation will hinge on remediation, not just discovery.

Related

• Nvidia Just Picked a Fight With Sequoia’s $11B Bet
• Claude Opus 4.7 Coding Score Breaks 87% Barrier
• AlphaGo’s Creator Just Made AI’s Boldest Bet Yet
• SoftBank’s $100B AI-Robotics Wager Just Got Real
• Zuckerberg’s Real Play With the Ari Humanoid Deal

Sources

1. CNBC — Anthropic CEO warns of cyber ‘moment of danger’ as AI exposes thousands of vulnerabilities (May 5, 2026)
2. Decrypt — ‘Moment of Danger’: Anthropic CEO Warns of Cyber Risk Window as AI Uncovers Software Flaws
3. Fortune — Jamie Dimon and Dario Amodei sidestep question about whether the AI cyber ‘freakout’ is warranted
4. PYMNTS — Anthropic CEO Predicts Firms Have 6 Months to Patch Software Vulnerabilities