EU Tests Anthropic's Mythos 5 Model as Security Concerns Prompt Fresh Safeguard Scrutiny

The European Union's cybersecurity agency has gained direct access to Anthropic's advanced Mythos 5 artificial intelligence model for testing following weeks of diplomatic negotiations. Brussels pursued access to the system amid mounting international scrutiny over its sophisticated hacking capabilities and potential security risks.

Following our constructive engagement with Anthropic, we can confirm that the EU's cybersecurity agency ENISA has been granted access to Mythos 5 and is now testing it, EU spokesman for digital affairs Thomas Regnier said on Thursday. The Greece-based agency previously received access to OpenAI models such as GPT-6-ASTRA, according to regulatory disclosures.

The evaluation by European regulators arrives concurrently with a comprehensive threat intelligence report published by Anthropic documenting multiple instances of malicious exploitation. The company stated that it successfully disrupted attempts to utilize its Claude models for biological weapons research and cyber espionage campaigns targeting Ukrainian officials. The report outlined five separate instances where researchers in unsupported regions utilized virtual private server infrastructure to plan avian influenza mammalian-adaptation experiments.

Rapid improvements in Anthropic's models have raised new risks across global digital networks, according to Jacob Klein, the company's head of threat intelligence. A year ago, let's say you wanted to optimize a drone or optimize the software on a missile, the models just wouldn't be as good at that task as they are now, Klein told Reuters. The findings also detailed operators in China, Russia, and Yemen leveraging the platform to develop software for conventional weaponry, munitions, and associated guidance systems.

Anthropic reported that state-backed actors increasingly rely on multi-agent AI frameworks rather than simple chatbot interactions to orchestrate complex cyber attacks. One targeted operation against Ukrainian government and diplomatic sectors utilized automated routines to detect security defenses and iteratively rewrite malware code until detection was evaded, tradecraft matching the Midnight Blizzard group linked to Russian intelligence.

Simultaneously, the artificial intelligence firm accused multiple Chinese laboratories of running large-scale distillation attacks to extract foundational capabilities. Operators linked to Alibaba allegedly generated over 151 million exchanges between May and July 2026 using thousands of fraudulent accounts to train domestic models like Qwen. Additional unauthorized routing was detected from DeepSeek and Moonshot, though representatives for the named entities did not immediately respond to requests for comment.

The convergence of advanced security vulnerabilities and commercial competition has intensified debate over the adequacy of voluntary industry guardrails. Speaking at the Global Fintech Fest 2026, Indian Finance Minister Nirmala Sitharaman questioned whether corporate safety frameworks are keeping pace with technological acceleration. Sitharaman highlighted the recent resignation of Anthropic pretraining researcher Jacob Coxon, who stepped down to warn about the unmitigated race toward self-improving superintelligence.

Again, just this week, public attention was drawn across global markets when a researcher resigned from one of the world's leading AI laboratories, Sitharaman said. Coxon's resignation note argued that commercial laboratories were gambling with human safety by prioritizing rapid capability scaling over alignment guarantees. Sitharaman emphasized that the governance of frontier technology cannot be left solely to corporate engineering teams, urging regulators and executive boards to enforce continuous oversight.

ENISA will continue its technical evaluation of Mythos 5 in the coming weeks to establish formal risk parameters for European digital infrastructure. Meanwhile, global policy bodies face mounting pressure to determine whether existing statutory frameworks can manage autonomous systems capable of recursive self-improvement. Anthropic has integrated its recent threat findings into its frontier model enforcement processes, though international regulators have yet to establish binding compliance standards for sovereign security evaluations.