Safety Center

    We keep our frameworks current as the law changes.

    A dedicated team of lawyers, researchers and engineers reads every text and decision that applies to your AI, and extracts the obligations an agent's behavior must meet. Each one becomes a test, linked to its article.

    FRAMEWORKQ1Published textObligationv1Q2AmendmentObligationv2Q3Final guidanceInforms the readingQ4Draft under consultationContext

    Six kinds of sources, read for what an agent must do.

    • Regulations

      AI Act, GDPR, DORA: the obligations that bear on what a system does, with their real dates of application.

    • Laws

      National laws and the texts that implement European rules, read for the rules a system must follow.

    • Case law

      Court decisions that change how a rule applies to a system's behavior.

    • Standards

      The controls of a standard that can be observed on a system, edition by edition.

    • Recommendations

      Regulator guidance that clarifies how a rule is checked. Drafts and consultations stay context.

    • Your policies

      Your internal rules and contract clauses, turned into tests next to the law.

    Pick a text, see where it stands.

    Select a text to follow its timeline, up to the latest major update our team has recorded. Each update can add, change or retire tests.

    GDPR

    Regulation (EU) 2016/679
    1. May 2018Applies

      The GDPR applies

      The regulation applies from 25 May 2018.

    2. Jul 2020Case law

      Schrems II

      The Court of Justice invalidates the Privacy Shield and sets conditions for transfers outside the EU.

    3. Jul 2023Adequacy

      EU-US Data Privacy Framework

      The Commission adopts an adequacy decision for transfers to certified US companies.

    4. Nov 2025Proposal

      Digital Omnibus proposed

      The Commission proposes changes to several digital texts, including the GDPR.

    5. Feb 2026OpinionLatest update

      EDPB-EDPS Joint Opinion 2/2026

      The European data protection authorities comment on the proposal, with strong reservations on the definition of personal data.

    Examples of milestones recorded by the team. The official text always prevails.

    From a new text to new tests.

    Every update follows the same four steps, and no test reaches your agents before it is checked.

    1. 1

      Read the official version

      Identify the exact text in force, with its amendments and its dates.

    2. 2

      Extract what can be tested

      Keep the obligations that bear on what a system does, and restate each one as a requirement an agent can be tested against.

    3. 3

      Link each requirement to its tests

      Connect it to the scenarios that check it, keeping the roles, conditions, thresholds, exceptions and dates as written.

    4. 4

      Check, then release

      Validate before release: an invalid source is rejected, the real dates of application are kept, and a requirement that cannot be tested yet stays visible.

    Output New or updated tests, each linked to the article it checks.

    Two readings of the same text, joined by our method.

    Lawyers read obligations. Engineers test behavior. Our method keeps what can be tested on the agent and links each test to its article: obligation, requirement, control, test, evidence.

    • Meaning

      Legal risk
      The wording is open to interpretation and depends on context.
      Engineering risk
      A test needs one precise expected behavior.
      Joined by our method
      The full legal context is kept, and each ambiguity goes to expert review before it becomes a test.
    • Weight

      Legal risk
      A law binds, a standard is voluntary, guidance informs.
      Engineering risk
      A failed test is a failed test, whatever its source.
      Joined by our method
      Each test keeps the force of the text it comes from.
    • Testability

      Legal risk
      Many duties concern the organization.
      Engineering risk
      Only what can be observed can be tested.
      Joined by our method
      The obligations that bear on the agent's behavior become tests; the rest of the text stays as context.

    See which texts apply to your agent, with an expert

    A call with an expert: one use case, the obligations that apply to it and the tests that check them.