Anthropic has given ENISA, the European Union’s cybersecurity agency, access to Claude Mythos, the model it has kept out of public release on security grounds. ENISA is the first European body to join Project Glasswing, the restricted programme built around the model which until now had been limited to roughly 40 mostly US-based organizations and the UK’s AI Security Institute. The Financial Times reported the decision on 1 June. An ENISA spokesperson said access had been offered but the terms were still being agreed.
Two days later Anthropic widened the program sharply. On 3 June it said it was extending access to roughly 150 more organizations across more than 15 countries, Sweden among them according to the Financial Times. The same expansion brought in NATO, the identity firm Okta and South Korea’s Samsung, SK Hynix and SK Telecom. The New York Stock Exchange and its parent, Intercontinental Exchange, said separately they were deploying Mythos across their exchanges and clearing houses.
What Mythos Actually Is
Mythos is a general-purpose model, not a tool built for attacks. Anthropic announced it on 7 April and said its coding and reasoning were strong enough at finding and exploiting software flaws that releasing it openly would be irresponsible. Instead the company set up Project Glasswing, a vetted group of operators and maintainers who use the model to find and patch vulnerabilities in critical software. Members are restricted to defensive use. Anthropic also committed up to $100 million in usage credits and $4 million in donations toward open-source security work.
The 10,000 Figure is Anthropic’s Own
Anthropic says Mythos has found more than 10,000 high and critical-severity zero-day vulnerabilities across every major operating system and browser including a 27-year-old bug in OpenBSD and a long-standing flaw in FFmpeg. The more detailed accounting it published is narrower. Of roughly 23,000 potential issues the model flagged across more than 1,000 open-source projects, external security firms had confirmed 1,726 at the time SecurityWeek reported with over 1,000 rated high or critical and 65 vendor advisories published.
The headline number is a vendor estimate and worth reading as one. Researchers at watchTowr told CNBC that comparable results can be reproduced by orchestrating publicly available models and Anthropic did not dispute to CNBC that earlier models could already find vulnerabilities. When the Linux Foundation ran Mythos against curl, one of the most scrutinized pieces of open-source software in the world, it surfaced a single low-severity bug. The UK’s AI Security Institute, which tested the model was more pointed about the risk, in controlled evaluations where Mythos was directed and given network access, it carried out multi-stage attacks and found and exploited vulnerabilities on its own, work that would take human professionals days.
Why Europe Pushed to Get In
The access fight was about exposure, not prestige. The European Central Bank convened euro-area banks to discuss the implications after learning Mythos had found vulnerabilities in financial software used widely across the eurozone. Finance ministers pressed for inclusion, and senior Commission officials flew to San Francisco to make the case in person. European Commission spokesperson Thomas Regnier confirmed to Dark Reading that the Commission had held several meetings with Anthropic about the model. The timing matters: the EU AI Act’s risk-classification and oversight provisions take effect in August 2026, and the bloc had no way to compel an American company to share the model before then.
Where This Leaves Swedish Operators
Neither Anthropic nor the Financial Times has named the Swedish organisations being brought into the programme, so treat country-level inclusion as exactly that. The practical question for Swedish essential and important entities is not whether they get Mythos. It is whether they can act on the volume of findings it and tools like it produce. Sweden’s Cybersäkerhetslagen, the national transposition of NIS2, has been in force since 15 January 2026, and Article 21 requires documented risk management, incident response, and supply-chain security that boards must approve and can be held personally liable for. Essential entities face fines of up to €10 million or 2% of global turnover.
The gap CNBC’s sources described is the one to close: vulnerabilities are being found in hours and patched in weeks. A model inside ENISA does not shorten your patch cycle. Map which of your systems run OpenBSD, FFmpeg, and the open-source components Mythos has been scanning, confirm you can ingest and triage vendor advisories at speed, and make sure the evidence trail your supervisory authority will ask for already exists. Regulations from MSB and PTS under the new law are expected during the first half of 2026.
References
- Claude Mythos Preview
- Anthropic to Give Mythos Access to EU Cybersecurity Agency ENISA
- Anthropic Negotiates EU Access to Mythos
- Anthropic to Open Mythos AI to the EU’s ENISA
- Anthropic Shares Mythos With 150 More Organizations
- Anthropic Expands Claude Mythos Access to India and 15 Other Countries
- Mythos Detected 23,000 Potential Vulnerabilities Across 1,000 OSS Projects
- Mythos Finds a curl Vulnerability
- NYSE Joins ICE in Deploying Anthropic’s Claude Mythos for Cybersecurity
- EU Cybersecurity Agency Gains Access to Anthropic Mythos
This post is also available in:
Svenska
Related News
Estonia Plans Official Digital IDs for AI Agents in Global First
Estonia is preparing to issue official digital identity credentials to AI agents, in what would be the first state-backed identity framework designed specifically for non-human…
July 9, 2026 
Five Eyes Warns AI Will Outpace Cyber Defences Within Months
The Five Eyes intelligence alliance has issued a joint warning that AI-powered cyberattacks are advancing faster than most organisations' defences can adapt, and that the…
June 24, 2026 