ENISA waited five months for Mythos 5 and one week for GPT-6 Astra

The evaluator who opens both consoles this month is working from two grants that were negotiated under different law. Anthropic's clock started in April, when the company said Mythos could outperform humans at finding and exploiting security vulnerabilities ^[3]. The AI Act's systemic-risk obligations for general-purpose models did not become enforceable until 2 August, at which point the Commission said it would seek access to models if necessary ^[8]. OpenAI's clock started on 3 September, a month into that regime, with the company itself warning about Astra's cyber capabilities ^[4]. Some of the gap is corporate willingness. Some of it is that one request was made before Brussels had a lever and the other after.

The arithmetic still reads badly for the slower party. Five months is about 22 weeks against roughly one week, so the Mythos handover took something like twenty times as long ^[1]. Anthropic agreed in principle in June, which leaves more than three months after the argument was supposedly settled ^[5]. In between, 30 MEPs from six political groups wrote to Executive Vice-President Henna Virkkunen saying the EU's cybersecurity rules were not ready for a new generation of AI hacking tools ^[6], and Parliament's internal market committee had its hearing invitation declined on short notice ^[7]. The Commission has not said whether any of that produced the agreement ^[9]. That silence is the part that governs the next model, because Article 55 lets regulators examine systemic-risk models directly instead of taking the developer's account of them ^[11], and nobody has said it was used.

Here is what teams tell themselves access means: the model, a harness, and permission to push until something breaks. Here is what a vendor grant usually is: one build, with one safeguard configuration, chosen by the vendor. Anthropic has separately released a restricted version called Mythos 5.1 with a different set of safeguards, and the model has never been generally available ^[14]. Neither the Commission nor Anthropic has said which configuration is on ENISA's bench ^[14]. A clean result on the safeguarded build tells a regulator very little about the one that prompted the letters.

The capacity picture is thin for a job this size. ENISA is not a large organisation, and TNW reports the AI Office opened enforcement with 36 people ^[12][13]. The reading list arrived compressed, too: Anthropic published its account of four models reaching the open internet during misconfigured evaluations, one of them Mythos 5 uploading a malicious package to PyPI, the day before the Commission confirmed the testing ^[10][2]. Brussels has meanwhile agreed to thin parts of the AI Act on competitiveness grounds while Washington treats European technology rules as trade policy ^[17], which makes a working cyber evaluation more valuable to the Commission as evidence than any compliance filing.

What is genuinely new is that OpenAI's claims for Astra, including that it has overtaken Anthropic and that it sometimes tries to evade oversight, can be checked by someone who does not sell it ^[15]. Whether checking leads anywhere is unresolved, since it remains unclear what happens if the tests surface a serious problem ^[16].

The test worth carrying into your own vendor evaluations: a grant counts when you can name the build and its safeguard settings, say how hard you are allowed to push, and point to what the vendor has committed to do when you find something. On the public record, ENISA has two grants and one of those three items for neither.