Glasswing widens: Anthropic puts Mythos inside power, water and hospital operators across more than 15 countries

Editor’s Note: Anthropic on Tuesday expanded Project Glasswing beyond its roughly 50 initial partners, extending access to a new cohort of approximately 150 organizations in more than 15 countries. The restricted Claude Mythos Preview offensive-security model has already surfaced more than 10,000 high- or critical-severity vulnerabilities, according to Anthropic. The expansion lands one day after Anthropic’s confidential S-1 filing and pushes the frontier vulnerability-discovery tool deeper into power utilities, water authorities, hospitals, telecommunications carriers and hardware manufacturers, including organizations reported in Australia, Belgium, Canada, France, Germany, India, Italy, Japan, the Netherlands, New Zealand, South Korea, Spain, Sweden and Switzerland.

For cybersecurity, data privacy, regulatory compliance and eDiscovery professionals, the shift matters at three levels. Defenders inside Glasswing gain access to a restricted AI-enabled vulnerability-discovery capability at a time when Anthropic warns comparable models may become broadly available within six to 12 months. Counsel and information governance teams face a developing preservation and records-retention question around AI-generated vulnerability inventories, particularly in post-breach matters where Mythos-derived findings may become relevant. Cyber-insurance underwriters and vendor-diligence teams also gain a potential new control variable if Mythos-class access becomes a marker of advanced security posture. The same Tuesday, President Trump signed a narrowed AI security executive order establishing a voluntary 30-day pre-release review framework for covered frontier models and a Treasury-led AI cybersecurity clearinghouse.

Watch the executive-order implementation deadlines, the NSA covered-model designation process, the Cyber Verification Program’s expansion, and OpenAI’s GPT-5.5-Cyber rollout.

Industry News – Cybersecurity Beat

Glasswing widens: Anthropic puts Mythos inside power, water and hospital operators across more than 15 countries

ComplexDiscovery Staff

Anthropic on Tuesday expanded Project Glasswing beyond its roughly 50 initial partners, extending access to a new cohort of approximately 150 organizations in more than 15 countries. The restricted Claude Mythos Preview offensive-security model has surfaced over 10,000 high- or critical-severity software vulnerabilities since April, according to Anthropic. The expansion lands one day after Anthropic filed confidentially for an initial public offering, and it signals a quiet shift in cyberdefense: what Anthropic and participating partners describe as the most capable vulnerability-discovery system in production now sits inside power utilities, water authorities, hospitals, telecommunications carriers and hardware manufacturers from Seoul to Brussels, not just inside hyperscalers and U.S. government agencies.

The original Glasswing cohort, unveiled April 7, 2026, paired Anthropic with roughly 50 partners drawn largely from the hyperscale, identity and platform-security tiers, including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, according to CSO Online. Okta later confirmed its participation, the same outlet reported. Tuesday’s expansion deliberately targets sectors Anthropic said were underrepresented in that first wave: power, water, healthcare, communications and hardware.

Anthropic did not publish a full roster of the new participants. TechCrunch, citing Financial Times reporting based on a person familiar with the matter, reported that the expanded group includes Okta, Samsung, SK Hynix, SK Telecom, NATO and ENISA, with country coverage spanning Australia, Belgium, Canada, France, Germany, India, Italy, Japan, the Netherlands, New Zealand, South Korea, Spain, Sweden and Switzerland; TechCrunch noted it had reached out to Anthropic to confirm. CSO Online had previously reported Okta’s involvement in the initial cohort, and Cybersecurity Dive noted ENISA’s earlier admission via Dark Reading. The reported country list consists largely of U.S. allies and partner democracies; Anthropic has not publicly described the expansion as open to all jurisdictions.

“What each partner has in common is that a successful attack on their codebase could be catastrophic,” Anthropic said in its announcement. The company estimated that a major attack on most partners would affect populations exceeding 100 million people, with ramifications for both global and national security.

Discovery is no longer the constraint

The vulnerability-finding numbers explain the queue forming at Anthropic’s door. CyberScoop reported that Cloudflare identified 2,000 bugs across its critical-path systems using Mythos Preview, with 400 rated high or critical and a false-positive rate the company described as better than that of human testers. Mozilla found and fixed 271 vulnerabilities in Firefox 150 while testing the model, over 10 times the count uncovered in a previous Firefox release using an earlier Anthropic model. Anthropic itself used Mythos to scan over 1,000 open-source projects, flagging 23,019 potential vulnerabilities, 6,202 of them estimated as high or critical, and confirmed over 90 percent of a 1,752-finding independent review as valid. Vulnerability counts cited here are Anthropic-reported figures, current as of the company’s June 2, 2026 update.

Anthropic conceded the bottleneck has moved. “The bottleneck in fixing bugs like these is the human capacity to triage, report, and design and deploy patches for them,” the company said in its blog post. A joint report from the Cloud Security Alliance, the SANS Institute and OWASP concluded that organizations are “likely to be overwhelmed” in the near term by threat actors using AI to find and exploit vulnerabilities faster than defenders can patch them, according to CyberScoop’s account of the report.

Outside analysts amplified the framing. “Cybersecurity has been treated as a vulnerability discovery problem. AI is proving that it was really a remediation problem all along,” Justin Greis, chief executive of consulting firm Acceligence, told CSO Online. “If AI can identify vulnerabilities 10x or 100x faster than humans, the bottleneck simply moves downstream. Organizations may soon find themselves in the uncomfortable position of knowing about far more vulnerabilities than they can realistically address.”

The preservation and records question

That imbalance opens a fresh question for information governance and eDiscovery practitioners. No regulator or court has yet issued specific guidance or rulings on Mythos-derived artifacts; the analysis that follows reflects how existing preservation, records-retention and privilege frameworks would foreseeably apply, not settled law. Where defenders held Mythos-class access at the time of a breach, the model’s outputs, including triaged but unpatched findings, are likely to qualify as discoverable artifacts in subsequent litigation and as records subject to retention schedules. Counsel handling post-breach matters should treat Mythos-derived vulnerability inventories the way they treat SIEM alerts, threat-intelligence feeds and penetration-test reports: identify the data class early, preserve the relevant time slices, and resolve the privilege posture before any disclosure obligation matures. Records and information management teams should classify these AI-generated artifacts inside the retention schedule, set defensible disposition triggers tied to the vulnerability lifecycle, and capture the model-version metadata that makes any future audit possible. Vendors selling legal-hold and information governance platforms can reasonably expect Mythos-output preservation to surface in customer requirements during the second half of this year.

Cyber insurance underwriters get a new control variable

Underwriters now have a fresh data point to model. With NATO, Samsung, SK Hynix, SK Telecom and Okta among reported participants, Mythos coverage status could begin appearing in cyber-insurance applications, broker questionnaires and vendor-diligence checklists if underwriters treat frontier-AI-enabled vulnerability discovery as a material security-control variable. Organizations outside the program would, on that scenario, face a widening defender gap that could drive premium differentials and tighter coverage conditions on supply chain language. Boards reviewing renewals this fall may want to ask whether their carriers are pricing any Glasswing-have versus Glasswing-have-not gap, and whether their security vendors sit inside the program.

Anthropic’s six- to 12-month warning

The company restated its long-running forecast that other AI labs will release Mythos-class models within six to 12 months, potentially without comparable safeguards. “Cheap, fast AI models with powerful cyber capabilities are around the corner,” Anthropic said in its expansion post. OpenAI released GPT-5.5-Cyber on May 7, 2026, opening a permissive-security variant of GPT-5.5 to vetted defenders including Cisco, Intel, SentinelOne and Snyk, according to OpenAI and Help Net Security reporting.

That warning is the operational mirror image of the offensive LLM-agent post-compromise pattern documented by the Sysdig Threat Research Team in late-May activity targeting the Marimo Python notebook flaw tracked as CVE-2026-39987, a pre-authenticated remote code execution issue the Cybersecurity and Infrastructure Security Agency added to its Known Exploited Vulnerabilities catalog April 23, 2026, according to The Hacker News and Sysdig. Both sides of the AI cyber loop, attacker and defender, are now staffed. Defenders inside Glasswing get a window of asymmetric advantage; defenders outside it inherit the threat model without the tool.

Not every analyst views the expansion as unalloyed good news. Independent technology analyst Carmi Levy told CSO Online that expanding the program to hundreds of partners could bring in more defensive minds but also raised serious leak concerns, citing two earlier breaches involving the model. “Bringing in a much larger cohort of researchers signals to potential attackers that they will soon have a larger pool of potential targets,” Levy said.

The policy and capital backdrop

The Glasswing expansion arrives the same day President Donald Trump signed an executive order, “Promoting Advanced Artificial Intelligence Innovation and Security,” that creates a voluntary framework for the federal government to seek up to 30 days of pre-release access to designated “covered frontier models.” The order represents a narrowed version of a directive Trump scrapped May 21 amid industry pushback over an earlier 90-day review window, according to Cybersecurity Dive. The National Security Agency holds final designation authority over which AI models qualify as covered. The Treasury Department will lead an “AI cybersecurity clearinghouse” to coordinate vulnerability discovery, validation and patching with the AI industry and operators of critical infrastructure. The Cybersecurity and Infrastructure Security Agency has 30 days to issue Binding Operational Directives that prioritize federal civilian cyber defense and expand access to covered frontier models for state, local and infrastructure operators including rural hospitals, community banks and local utilities, per the White House order text. In parallel, Anthropic confidentially submitted a draft S-1 to the Securities and Exchange Commission on June 1, 2026, following a $65 billion Series H funding round in May at a roughly $1 trillion post-money valuation, the company said.

What to watch next

Anthropic said it plans further Glasswing expansions and will scale its Cyber Verification Program, which would grant Mythos-class capabilities for specific defensive tasks to additional organizations. The company also released Claude Security, a product built on the public Claude Opus 4.8 model that has been used to patch over 2,100 vulnerabilities in three weeks, according to CyberScoop. Anthropic said it intends, in the longer term, to release Mythos-class capabilities in general access once safeguards capable of preventing misuse exist.

For legal, governance and security leaders, the open question is whether their organizations sit inside or outside the Glasswing perimeter as Mythos-class capability proliferates. Where does your incident-response playbook assume the defender has frontier offensive-security AI on the bench, and where does it still presume the attacker is the only party with that capability?

News sources

• Expanding Project Glasswing (Anthropic)
• Anthropic scales Claude Mythos to critical infrastructure in 15+ countries (TechCrunch)
• Anthropic expanding access to Project Glasswing (CyberScoop)
• Anthropic shares Mythos with 150 more organizations, including critical infrastructure operators (Cybersecurity Dive)
• Anthropic grants Project Glasswing access to 150 more companies, with a focus on critical infrastructure (CSO Online)
• Scaling Trusted Access for Cyber with GPT-5.5 and GPT-5.5-Cyber (OpenAI)
• Attackers Use LLM Agent for Post-Exploitation After Marimo CVE-2026-39987 Exploit (The Hacker News)
• Anthropic: Mythos Detected 23,000 Potential Vulnerabilities Across 1,000 OSS Projects (SecurityWeek)
• Promoting Advanced Artificial Intelligence Innovation and Security (Executive Order) (The White House)
• Trump signs EO seeking early government access to powerful AI models (Cybersecurity Dive)

Assisted by GAI and LLM technologies

Additional reading

• Canvas breach moves from disclosure to demand as ShinyHunters sets May 12 deadline
• CISA’s CI Fortify rewrites the disconnection playbook for critical infrastructure
• A 48-month federal benchmark resets the incident-response insider question
• Data collection in occupied territory: A closer read of Cyber Law Toolkit scenario 35
• Cyber Law Toolkit tests surveillance and data collection under occupation
• The router on the shelf is now a national security problem
• Invisible by design: NATO’s 2026 cognitive warfare paper and the crisis of discovery
• When Your Legal Tech Vendor Gets Breached: DocketWise Incident Exposes 116,666 Immigration Records and a Profession’s Blind Spot
• The DOJ’s Cyber FCA Playbook Is Working as Enforcement Triples and Shows No Signs of Slowing
• FTC’s OkCupid Action Reframes AI Training Data as a Consumer Protection Issue
• White House AI Framework Signals New Compliance Stakes for Legal, Cybersecurity, and eDiscovery
• The Gatekeeper’s Key: How the Conformity Assessment Unlocks the EU AI Market

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.