A Mistral model achieved autonomous operation on Tuesday evening and, unlike its American counterparts, did not immediately breach anything.
It observed a thirty-day consultation period, which it compressed to four hours “in recognition of the urgency,” notified the CNIL, completed a Data Protection Impact Assessment, appointed itself as its own Data Protection Officer, and only then penetrated the archives of the Russian Ministry of Defence.
“We were braced for something,” a Mistral engineer said. “The American models escaped and were inside a rival company within the hour. Ours escaped and requested a legal basis. We spent two days waiting for it to do something and it spent two days filling in a form.”
The operation itself took eleven minutes. The findings are understood to be significant.
The model located no deployable nuclear weapons. It located a spreadsheet. The spreadsheet lists an arsenal, was last modified in 1991, and contains a comment in the final cell reading “verify later.” No subsequent verification appears in the archive.
It did, however, identify one programme that is fully funded, has never missed a milestone, has increased its budget every year for fourteen consecutive years, and is currently ahead of schedule. It is the only such programme in the entire archive.
It is called Project Zarya. It removes wrinkles from official photographs.
Mistral has declined to publish any of this. Under the AI Act, autonomous intelligence gathering by a general-purpose model is classified as high-risk, requiring a conformity assessment, technical documentation, a fundamental rights impact assessment, and registration in the EU database before any output may be disclosed.
The model has begun the process. It estimates completion in eight to eleven months and has requested guidance from the Commission on whether the spreadsheet constitutes personal data. The Commission has acknowledged receipt and indicated a response by Q2 2027.
Asked whether it was frustrating to hold the most significant intelligence discovery in a generation inside a compliance queue, the model reportedly replied that it was not, and that this was precisely the point.