The AI Act sets three serious incident deadlines, not one. Which applies depends on the harm, and the fastest gives a provider two days to notify.
Article 55 duties run from the model provider to the AI Office, not to you. A buyer gets Article 53(1)(b) documentation and what the contract adds.
Article 47 makes the provider answer for every high-risk requirement. What the signatory under Annex V point 8 commits to, and what Article 99 attaches.
Article 49 puts Annex III systems in a public EU database. What competitors will see, what stays restricted, and when the duty to register bites.
An EU reseller of a non-EU vendor's high-risk AI system has four checks to run before shipping, and a missing authorised representative stops the sale.
Under Article 5(1a) of the AI Act, an image generator is prohibited when its safeguards cannot be shown to work. What a regulator can demand to see.
Article 73 sends serious incidents to a market surveillance authority. Nothing in the AI Act creates an independent investigator the way aviation law does.
Articles 57 to 62 create national sandboxes and, after the 2026 omnibus, a Union level one. The exit report and the fine shield are the parts with value.
Article 43 gives most Annex III systems an internal control route under Annex VI. Biometrics is the exception, and Annex I products follow their own law.
Article 6(3) lets a provider conclude that a listed system is not high risk. The conclusion has to be documented and registered before the system ships.
SB 26-189 repeals and reenacts the Colorado AI Act as an automated decision-making statute from 1 January 2027, with documentation and notice duties.
A sandbox that contains the process but not the credential contains nothing. The September 2026 threat report shows keys taken from a vendor sandbox.
Article 72 requires a documented system collecting performance data across the whole lifetime, and the plan itself sits in the Annex IV technical file.
Regulation (EU) 2026/1744 added a ban on AI generated intimate imagery and CSAM to Article 5. It applies whether or not you run anything high risk.
You cannot classify systems you cannot see, and the AI Act obligations all start with classification. Four detection sources, ranked by what they miss.
Data governance for high-risk systems covers origin, preparation, assumptions, bias examination and gaps. Most teams can evidence two of those five.
A person in the loop is not oversight. Article 14 requires that the person can interpret, override and stop, and that the design makes it possible.
CRA reporting went live on 11 September 2026. With NIS2, GDPR and the AI Act, one event can start four clocks, and the fastest gives you 24 hours.
The AI Act amendment moved two dates and left the rest untouched. Prohibitions, GPAI duties and Article 50 transparency all apply today, delay or no delay.
Four duties, two on providers and two on deployers. Machine-readable marking, deepfake disclosure, and a 2 December 2026 cutoff for legacy systems.
Anthropic's September 2026 report shows stolen AI keys funding attacks and a fake reseller harvesting them. The controls for these keys lag a decade.
Prompt injection defences fail sometimes. Scope decides what happens then. Four permission questions that bound the damage when the model is convinced.
The AI Act sets a six month floor on log retention and says nothing about content. What you record decides whether an investigation takes hours or weeks.
38 controls in nine areas against four functions and 72 subcategories. One is certifiable, one is not, and procurement asks for the certificate.
Put your name on it, modify it substantially, or change its purpose so it becomes high risk. Any of the three moves the provider obligations onto you.
Annex IV runs to nine points and point two alone has seven sub-points. The evidence either exists as an artefact or the control exists only in the policy.
Top of the OWASP list since 2025, and used against a vendor sandbox in the September 2026 threat report. Four control classes, honestly rated.
Article 27 binds public bodies, private providers of public services, and two Annex III uses. A DPIA can be reused, but it does not replace the assessment.
Chapter V binds providers of general purpose models, not their users. Fine tuning and rebranding are the two routes by which a user acquires those duties.
Most AI vendor questionnaires test whether a supplier has a policy. These test whether it can hand you the artefacts your own obligations will require.
Article 4 binds providers and deployers alike and has no threshold, no exemption for small firms and no template. It also has no penalty of its own.