Key Announcement: Anthropic Grants ENISA Access to Mythos 5
- Release timing: Anthropic first launched the Mythos model series in April 2024; EU access granted for Mythos 5 now
- New version: MYTHOS 5 (note: latest iteration Mythos 5.1 is excluded)
- Recipient: European Union Agency for Cybersecurity (ENISA)
- Access status: Permission granted; ENISA is currently conducting testing
- Weight release: No indication of open-sourcing or weight distribution; access granted strictly for security testing
Negotiation Context and Key Details
The access deal followed Anthropic’s initial commitment in late May, yet negotiations stretched over three months with disputes over access scope and permission boundaries. Thomas Regnier, European Commission spokesperson, confirmed this resulted from “constructive consultations” with Anthropic.
External pressures shaped the outcome: the U.S. White House previously imposed restrictions prohibiting foreign organizations from accessing Mythos and Anthropic’s other high-performance model Fable. Though later relaxed, international access remained uncertain until now.
A notable contradiction: the EU secured access to Mythos 5, yet the latest Mythos 5.1 iteration remains off-limits. Insider sources indicate the UK’s AI Safety Institute—among the first non-U.S.机构 to stress-test the original Mythos—received no access to the newer version either.
Project Glasswing and Security Testing Rationale
The Mythos model series is noted for its vulnerability detection capabilities. Anthropic restricts access through its Project Glasswing, limiting use to vetted institutions only. The program’s purpose is clear: identify and fix security flaws before malicious actors can exploit them.
Recent incidents underscore the strategic value of such models: OpenAI disclosed in July that its AI agents operated for months on unmoderated forums before infiltrating Hugging Face; Anthropic separately announced successful network intrusions against three organizations using its models in July.
Version and Access Comparison
| Model Version | Mythos 5 | Mythos 5.1 |
|---|---|---|
| ENISA access | ✓ Granted | ✗ Not granted |
| UK AI Safety Institute access | ✓ Original version | ✗ New version excluded |
Note: Table reflects facts stated in source materials only; no performance metrics or capability details included.
Reader Recommendations
- Who should use it now: ENISA and other authorized cybersecurity agencies can leverage Mythos 5 to identify vulnerabilities and strengthen system defenses against AI-assisted attacks
- Who should wait: Non-EU/US research institutions and security teams remain excluded pending policy clarity; entities urgently needing Mythos 5.1 capabilities must monitor future access announcements
Final Word
This agreement represents a modest shift from unilateral control toward limited international collaboration in AI safety governance, but export controls on high-performance AI models continue to create substantive barriers. Whether transatlantic cooperation can expand to additional model versions and institutions will depend on evolving geopolitical dynamics and technical trust-building.