678,000 French filers and one 9.4: the week's patch-and-notify work, with numbers attached

The value of the DGFiP set to whoever holds it is in the pairing. A withholding rate next to a property record and an address describes a household's finances precisely enough to script a phone call ^[2]. DGFiP says online accounts and passwords were untouched ^[3], which takes account takeover off the table and leaves impersonation, and impersonation is the harm you cannot remediate centrally. A password reset fixes a credential. It does not make a filer's declared income less accurate in a fraudster's hands ^[9].

The 678,000 is also a confirmed count on an open investigation ^[1][3]. Read it as the number DGFiP can currently stand behind rather than the number the incident produced ^[10]. Notification programmes get sized off that first figure, and they are expensive to re-run.

The GitLab item is the cleaner piece of work, because the fix is schedulable. An unauthenticated remote attacker able to modify or delete public projects and user data is an integrity and availability problem before it is a confidentiality one ^[4]. Nothing has to leave the instance for the damage to land; a deleted public project is missing for everyone who was pulling from it, and the second fix in the same batch, a high-severity GraphQL CSRF issue, sits behind it in the queue ^[5]. Both carry identifiers and severity ratings, which is the whole reason a maintenance window can be argued for on Monday morning.

Set that against the roundup's enterprise AI entry. Guild Group's Mohammad Arif names six risk categories: shadow AI, data leakage, insecure integrations, AI supply-chain attacks, prompt injection and AI-powered phishing ^[8]. None of that is wrong. It also has no incident behind it, no affected population, and no score, which means it cannot be ranked against the GitLab patch or the French notification effort by any method other than assertion. Security teams allocate hours against the second kind of item and get asked about the first kind in board papers.

The concrete side of the week is not uniformly better, either. Oz Hair and Beauty confirmed which fields an unauthorised party reached, and confirmed which ones it did not, while leaving the number of affected customers undisclosed ^[7]. So of the week's two customer-data confirmations, one produced a population and one produced only a field list ^[11]. Field lists tell you the fraud pattern. Populations tell you the cost.

Ukraine's ARMA supplies the week's one piece of inference dressed as reporting, and it is candid about being inference: the agency said a suspected attack shortly before the deadline to select a manager for assets linked to sanctioned oligarch Mikhail Fridman, taken with earlier cyber activity and increased information pressure, could point to a coordinated attempt to disrupt the tender ^[6]. That is a victim reading timing as motive, which is reasonable for a body running a contested procurement and still short of attribution.

Two items this week can be closed out with a version bump and a mailing list. The rest is posture.