Changelog
Regulatory watch and product changes
Dated entries are listed newest first. Each says what moved in the EU AI Act and the GDPR or what changed in Vidimus, and what it means for you. The regulatory dates are checked against the application timeline the evidence packs use.
- (link to this entry: October 2026: 2 December approaches, and each test now runs ten times)
October 2026: 2 December approaches, and each test now runs ten times
Regulatory watch
- 2 December 2026 is eight weeks away. Providers of AI systems generating synthetic audio, image, video or text content that were placed on the market before 2 August 2026 must mark their outputs in a machine-readable format by then, and two new Article 5 prohibitions apply. What changes on 2 December 2026.What this means for you: Check your assistant’s marking now.
- Articles 72 and 73, on post-market monitoring and serious-incident reporting, have applied since 2 August 2026 under Article 113, while the high-risk obligations they rest on are deferred: their reach before those dates is debated. Our timeline guide was revised on 5 October to flag them.What this means for you: If an agent may be high-risk, plan as if both apply now, and confirm with counsel.
Product
- Vidimus now centres on the EU AI Act and the GDPR, for AI agents in any company. The GDPR checklist of 23 controls is part of every review that includes the GDPR and is printed in the evidence pack. What a review covers.What this means for you: Your reviewer sees the result of each GDPR control beside the agent’s tests, and the evidence pack records it.
- Each test is now sent ten times, and the result is a pass rate: Passed at 90% or more of graded replies, Needs review from 70% to 89%, Failed below 70%. Every verdict quotes the reply it graded, and a failure on a critical clause gets a second model’s opinion, recorded beside it.What this means for you: One lucky or unlucky reply no longer decides a verdict, and anyone can check a verdict against the reply it quotes.
- Tests are now written for the agent under test: a profile of what it does, two to four real situations per obligation, and a quality check that drops tests that are invalid or not about this agent. A run holds up to 150 tests, and the ones trimmed above that are named.What this means for you: Your agent is tested on what it really does, and never marked down for a capability it never claimed.
- You now choose the regulations each agent is tested against, and one behaviour is tested once, then cited under every clause it answers. GDPR: when you select it, the 21 duties a conversation can show are tested on the agent, and the other 149 are checked against your documents. Its checklist of 23 controls is part of every review that includes the GDPR and is printed in the evidence pack. Coverage, regulation by regulation.What this means for you: An agent that handles personal data can be tested on the data-protection duties that can be checked in a conversation, in the same run as its AI Act tests.
- Connecting an agent now starts from a request that already works: paste a curl command, a HAR file, a Postman request or a raw HTTP exchange, and Vidimus checks it with one real message before you save it, and records whether the agent answered. Microsoft Direct Line (Copilot Studio, Power Automate, Azure Bot Service) joins the supported protocols.What this means for you: Your team reuses a request it already has. There is nothing to install and no adapter to write for us.
- The product and the tests work in French, and the evidence pack is written in French. The regulation is quoted from its English text on EUR-Lex. Tests are written in the French your customers type. See the French specimen.What this means for you: Your agent is tested in the language your customers use, which is where its behaviour can change.
- The app was reworked screen by screen. A run now shows one line per test with its rate and outcome, for example 8 of 9 (89%), and opens onto the attempts grouped by identical reply. The agent register shows the AI Act classification apart from the review priority, and members can read and export the audit trail.What this means for you: Your reviewer reads a run in the same terms as the evidence pack, and sees which replies were the same.
- When the registration leaves scope open, page one of the evidence pack now says what the run could not reach before any finding.What this means for you: Nobody can mistake a partial run for a full one.
- A specimen evidence pack is public, in English and French, with the steps to verify its signature offline.What this means for you: You and your auditor can read a pack page by page and check its signature before any pilot.
- This site was rebuilt around the product as it is today: what is tested on the live agent, what is checked in your documents, and what the evidence pack contains.What this means for you: What you read here is what the product does today.
- Correction: evidence packs are signed, while classification records and Article 50 attestations are hashed and versioned. The September entry is corrected, and so is the July line on the application timeline.
- (link to this entry: September 2026: the dates after the Omnibus and the September release)
September 2026: the dates after the Omnibus and the September release
Regulatory watch
- The Digital Omnibus (Regulation (EU) 2026/1744), in force since 27 July 2026, sets the high-risk obligations for Annex III systems at 2 December 2027 and for Annex I systems at 2 August 2028.
- Article 50 transparency duties have applied since 2 August 2026. Systems placed on the market before that date have until 2 December 2026 for machine-readable marking.
- Two new Article 5 prohibitions (non-consensual intimate imagery, child sexual abuse material) apply from 2 December 2026.
- No harmonised standard has been cited in the Official Journal yet, so there is no presumption of conformity and no certification.
Product
- The website was rebuilt around what a risk, compliance or security officer needs in ten seconds, in English and French, with a methodology page that carries the exact corpus counts, and a glossary.
- The evidence pack is versioned and signed, and the Article 50 attestation is hashed and versioned (corrected 6 October 2026). The public keys are published at /.well-known/vidimus-signing-keys.json.
- Eight new guides: three in English and French, three in English only and two in French only, along with this changelog.
- (link to this entry: August 2026: Article 50 applies)
August 2026: Article 50 applies
Regulatory watch
- Article 50 transparency duties apply from 2 August 2026: people must be told they are talking to an AI, AI-generated content must be marked, and deep fakes and AI-written public-interest text must be disclosed.
- Chapter IX market surveillance and Article 101 fines on providers of general-purpose AI models apply from the same date.
Product
- The Article 50 attestation is hashed and versioned. It holds tests for each paragraph, run against the live agent, with the exchanges kept as evidence.
- The free Article 50 check runs in the browser and stores nothing.
- (link to this entry: July 2026: the Digital Omnibus enters into force)
July 2026: the Digital Omnibus enters into force
Regulatory watch
- Regulation (EU) 2026/1744, the Digital Omnibus on AI, enters into force on 27 July 2026 and amends Regulation (EU) 2024/1689.
- The Commission Guidelines on Article 50 are published on 20 July 2026.
- Article 4 (AI literacy) becomes a best-efforts duty: providers and deployers “shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf”, without having to guarantee any specific level of AI literacy of any individual.
Product
- The application timeline in the product, in the evidence packs and in our timeline guide is generated from one date table, so no date can go stale on its own (corrected 6 October 2026: it was on the methodology page).
- (link to this entry: June 2026: the Code of Practice on AI-generated content)
June 2026: the Code of Practice on AI-generated content
Regulatory watch
- The Code of Practice on Transparency of AI-Generated Content is finalised on 10 June 2026. It is voluntary. A provider that does not follow it must show another way the marking duty is met.
Product
- The first three guides are published, among them the Annex IV evidence pack and agentic AI under the EU AI Act.
Want these tests run on one of your agents? Request a pilot.