Breaking

Anthropic Widens Project Glasswing to 150 Orgs Across 15 Countries

June 14, 2026 at 02:08 EDT

Security
Foundation Models
AI Agents

On June 2, 2026, Anthropic said it is expanding its defensive cybersecurity initiative "Project Glasswing," opening access to its unreleased frontier model "Claude Mythos Preview" to roughly 150 additional organizations based in more than 15 countries. Combined with the initial roughly 50 organizations, the move continues and broadens a gated program focused on finding and fixing software vulnerabilities, newly covering organizations in the power, water, healthcare, telecom and hardware sectors (announcement).

June 2, 2026 · Anthropic

Project Glasswing Scales Up: a Frontier AI Aimed at Defenders, Not Attackers

Anthropic extends gated access to its unreleased "Claude Mythos Preview" model to ~150 new organizations across 15+ countries — broadening its defensive cybersecurity coalition into power, water, healthcare, telecom and hardware.

Coalition growth

From ~50 vetted partners at the April 7 launch to roughly 200 — a ~4× expansion of participating organizations.

~50

at launch

~200

after expansion

10,000+

high/critical vulnerabilities found across partners

15+

countries now represented among partners

$100M

in usage credits, plus $4M donated to OSS security groups

Open-source scan results

Across 1,000+ open-source projects, 6,202 of 23,019 flagged issues were rated high or critical severity — about 27%.

6,202

16,817 other

high / critical 23,019 total flagged

The case for it

AI now rivals skilled humans at finding vulnerabilities — gating it to vetted defenders shifts the initiative from attackers to defenders and lets fixes ship before exploits.

The open questions

The volume of findings can bottleneck verification and remediation, the model ships without safety mitigations, and dual-use risk plus gated access fuel ongoing debate.

How access works

Gated to vetted partners → Scan · patch · pre-release checks → Share findings industry-wide

Claude Mythos Preview will not be released publicly; a similar, safeguarded model is planned for the future.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.