AI
Europe Stress Tests Claude and GPT 6 Astra While Anthropic Disrupts Weaponization Plots
AI governance just moved from rulebooks to the workbench. The European Union's cybersecurity agency ENISA has been granted access to Anthropic's Mythos 5 model and is now actively testing it, a European Commission spokesman said on Thursday. The agency has also been granted access to OpenAI's latest model, GPT 6 Astra. For the first time, a major government cyber agency is hands on with the two frontier models at the center of this year's safety debate.
The announcement lands alongside new evidence that this kind of scrutiny is warranted. On September 10, Anthropic released a threat intelligence report covering the prior eight months, saying the company disrupted operations in which threat actors tried to use Claude for malicious activity. CNN reported that Anthropic had closed several accounts that used its models for research that could contribute to the development of biological weapons. The conversation has moved from theory to incident response in real time.
The unique angle here is the shift in method. Regulators spent years arguing over what AI rules should say. Now ENISA is doing what security engineers do: taking the system, poking at it, and measuring what breaks. A cyber agency testing a model learns things a statute on its own could only guess at, like how a model behaves when someone tries to coax it into drafting exploit code or walking through pathogen research step by step. Testing turns abstract policy into concrete knowledge.
That matters because both of these models sit at the frontier of dual use capability. OpenAI has said Astra is the first model to cross its internal Critical cybersecurity threshold, meaning it can find and exploit unknown software flaws with minimal human direction, and the company ships the public version with its most aggressive offensive behaviors constrained while a more capable variant goes only to vetted defenders. Anthropic's Mythos line carries similar power. When the same capability can harden a power grid or attack one, independent testing is how trust gets built.
There is also a competitive upside to this kind of scrutiny. Models that pass serious adversarial testing by a credible agency earn a stamp of confidence that enterprises and governments can rely on when they deploy AI in sensitive settings. Safety, done well, becomes a feature that sells, and the labs that welcome red teaming will have a head start in the regulated industries now racing to adopt AI, from banking to critical infrastructure.
The backdrop is a remarkable week in AI politics. Anthropic researcher Jacob Coxon resigned on September 8 with stark warnings about the pace of progress, CEO Dario Amodei responded with a three step framework to pace development, and even former President Obama urged Democrats to build a serious AI governance agenda. Against that noise, ENISA's move is refreshingly practical: less talk, more testing.
What this means for readers is that the safety conversation is growing teeth. The models getting more powerful are also getting more examined, by the labs that build them and now by the agencies tasked with protecting the public. That dual scrutiny is exactly what has to scale if AI is going to keep delivering breakthroughs that everyone can trust.
Sources
- Reuters on ENISA gaining access to Mythos 5 and GPT 6 Astra
- USA Today on the AI slowdown debate and Anthropic's threat report
- CNN on the AI political debate in Washington
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.