Skip to main content
Client access Talk to us

Latest edition

Blog · 267 analyses

Regulation, AI and cybersecurity, with sources in view.

Context for better decisions.

Analysis on AI, cybersecurity, regulation, governance and digital systems. With sources, practical consequences and one clear next action.

Showing

Artificial intelligence

49–54 of 54

Current knowledge, organised by topic, date and practical consequence.

Operator controls a stop lever between decision modules while a reviewer follows the evidence
Oversight exists only when the person understands the context, has authority, and can intervene before the effect.AI-generated image under Velqory editorial direction.
Artificial intelligence13 min read

Human oversight of AI: the gate needs authority, time, and proof

High-risk requirements moved to December 2027 — or August 2028 where AI is embedded in certain products. The additional time should be used to test human intervention, not to add an approval button at the end of a process.

Read analysis
Two independent evaluators compare generic modules, calibration instruments, and evidence folders
An evaluation is credible only when independence, scenario, configuration, results, and limitations can be reconstructed.AI-generated image under Velqory editorial direction.
Artificial intelligence11 min read

External AI evaluation: seven proofs before accepting a report

The AI Office is developing independence and qualification requirements for external evaluators of general-purpose models with systemic risk. Even outside that perimeter, buyers need to distinguish reproducible testing from a commercial seal.

Read analysis
Technician separates untrusted documents from instructions before a generic decision device
Retrieved content is untrusted data; authorisation and validation must exist outside the model.AI-generated image under Velqory editorial direction.
Artificial intelligence12 min read

Prompt injection: test RAG and copilots beyond the input filter

NIST and OWASP treat prompt injection as an attack class that can arrive through direct requests or external content. In systems with retrieval and tools, asking the model to ignore malicious instructions is not a security boundary.

Read analysis
Conceptual installation of connected modules and a system opened for inspection
Useful enforcement readiness starts with knowing the organisation's role and the evidence it can produce.AI-generated image under Velqory editorial direction.
Artificial intelligence12 min read

EU AI Act on 2 August: enforcement begins for general-purpose AI models

Tomorrow the Commission gains effective powers over applicable obligations for providers of the most advanced models. The AI Omnibus delayed high-risk rules, not the entire regulation: this guide separates roles, dates and evidence.

Read analysis

How we publish

Each analysis links to the official source and says what changes in practice.

  1. 01Sources in view

    Every analysis links the official documentation it relies on.

  2. 02Operational context

    We translate change into decisions, owners and evidence.

  3. 03Continuous review

    We update the analysis when facts or guidance change.

  4. 04Identified imagery

    We disclose generated photography and credit documentary photography.

From context to execution

Is there an obligation or a system that needs to become operational?

We map the problem, write the scope and build the solution with acceptance criteria.