EU Mythos access report leaves model version and operating status unclear

Bloomberg News reported on 10 September that Anthropic has provided the European Union with access to its restricted Claude Mythos model. If confirmed, the development would move an effort previously described as under negotiation into a new phase. No located public statement from Anthropic, the European Union Agency for Cybersecurity, known as ENISA, or the European Commission independently confirms the reported arrangement or says access is operational.

The distinction matters because the available evidence does not establish which EU institution received Mythos, when the arrangement was approved or activated, or whether authorised users can work with the model directly. The exact model version is also unknown.

Bloomberg’s earlier reporting provides the clearest indication of the intended recipient. On 1 June, the publication reported that Anthropic was preparing to let ENISA join Project Glasswing. That report concerned prospective participation and did not demonstrate that the agency had begun using Mythos. Bloomberg’s latest account presents the effort as having progressed from a planned offer to reported access, although the recipient has not been independently confirmed in connection with the September development.

Any institutional arrangement should not be confused with general European availability. Access for ENISA, if it is the recipient, would not by itself give EU member states, banks, companies, critical-infrastructure operators or the public direct use of Mythos. The scope of permitted users and the systems they may be allowed to assess remain undisclosed.

Five months of restricted access

Anthropic unveiled Claude Mythos Preview on 7 April and said the model would not be generally available. The company restricted external access because of what it described as advanced dual-use cybersecurity capabilities. In its later material, Anthropic said Mythos-level capabilities in cybersecurity and biology required access to be limited to vetted organisations through trusted-access programmes.

European interest became public in May. Bloomberg reported on 4 May that EU officials had confirmed talks with Anthropic about using Mythos to test European companies and banks for vulnerabilities. The subsequent June report identifying ENISA as the prospective institutional recipient narrowed the immediate proposal, but it did not establish direct access for the companies or financial institutions that might ultimately be assessed.

A broader US policy constraint followed. Anthropic said a US government directive issued on 12 June required it to suspend access involving foreign nationals to both Fable 5 and Mythos 5. On 1 July, the company said those restrictions had been lifted and that Mythos 5 access had been restored to a set of US organisations.

Anthropic introduced Mythos 5.1 on 1 September while continuing to describe the model as restricted. Its public Mythos page said at that time that the company was able to provide access only to a set of US organisations. A European Commission press briefing the following day included questions about efforts to secure European access to Mythos 5.1.

Bloomberg’s report of EU access arrived nine days after Anthropic’s US-only availability statement. No public explanation reconciling those positions was located. Anthropic’s page may not reflect a subsequent arrangement, but no primary update confirming expanded European availability was found by the 10 September evidence cutoff.

Delivery and safeguards remain unclear

The discrepancy matters because access can describe several materially different arrangements. The public information does not show whether an EU body received direct model use, supervised testing, API credentials or access only to findings produced by Anthropic. There is no evidence that model weights were transferred or that unrestricted access was granted.

The contractual and technical controls surrounding the reported arrangement are similarly unknown. Anthropic, ENISA and the Commission have not publicly detailed hosting, monitoring, data residency, retention, geographic restrictions, permitted use cases or the systems in scope. It is also unknown whether the arrangement required US government authorisation and, if it did, whether that approval was granted.

These gaps are particularly relevant for a model whose developer has made controlled access part of its safety position. The reported EU arrangement could expand defensive evaluation capacity inside an EU institution, but its practical significance depends on who can use Mythos, under what supervision and for which targets. Without those details, the development cannot yet be treated as wider European commercial access or a transfer of independent sovereign capability.

There are also no confirmed operational results. No public evidence located for this report shows that an EU institution has used Mythos, discovered vulnerabilities, initiated remediation or produced findings for European banks, companies or infrastructure operators. No service disruption or data exposure resulting from EU use has been identified.

A limited precedent beyond Europe

The access question has relevance outside the EU because restricted foreign cyber-AI models can raise both security and sovereign-access considerations. Australia has no identified involvement in the reported arrangement, and no access to Mythos has been established for the Australian Signals Directorate, the Australian Cyber Security Centre, local researchers or Australian critical-infrastructure operators.

The European case nevertheless provides a limited policy parallel for whether those organisations could obtain comparable vetted access in future. For now, the central unanswered question is whether Bloomberg’s report reflects an approved arrangement, a live operational capability or another tightly supervised form of participation. Primary confirmation from Anthropic or the relevant EU institution is still needed to resolve that distinction.

Restricted model access puts scope, supervision and permission at the centre of responsible security evaluation. Examine security testing foundations with The Hack Academy’s online cybersecurity training programme, covering listed subjects such as cybersecurity fundamentals and penetration testing through self-paced learning and practical exercises alongside theory; practise only with authorisation and in controlled environments.