More from the AccuroAI blog.

Prompt injection is first on OWASP's Top 10 for LLM Applications. This guide covers direct and indirect attacks, the four surfaces they use against agents, the four defense layers that bound the damage, and what Germany's BSI, ETSI EN 304 223 and the UK code of practice say.

One rule commences on 13 November 2026. It is titled “Registration and obligations of Consent Manager”, and the whole one-year tranche across both Gazette instruments reaches no ordinary enterprise at all. What you owe arrives in May 2027, under a law that never says artificial intelligence, and whose only mention of algorithms currently applies to nobody.

Singapore's Model AI Governance Framework for Agentic AI, version 1.5, is voluntary guidance from IMDA, and it is specific enough to audit against. We map each part to a control and to the evidence an auditor would accept, with the Cyber Security Agency's agentic addendum and the PDPA alongside.

Four state AI laws took effect on 1 January 2026 and a fifth arrives in 2027. For most enterprises three of the five impose no direct obligation, and the question that sorts them is whether you develop AI or deploy it. Colorado is also the law most published guidance still describes incorrectly.

From 10 December 2026, an Australian privacy policy must list the kinds of personal information and the kinds of decisions involved wherever a computer program makes, or substantially shapes, a decision that could significantly affect a person. The OAIC's guidance of 30 September 2026 counts generative AI as a computer program and says advisory outputs can be in scope. You cannot disclose decisions you have not found, so the work starts with an inventory.

When an employee pastes customer data into a chatbot nobody approved, the provider has no Article 28 contract, the processing is missing from your Article 30 record, and the Dutch DPA treats the paste itself as a data breach. Agents add Article 22 once software acts on people. This guide maps each AI activity to the GDPR article it triggers and the evidence a supervisory authority will ask for.

Most summaries call Section 552.105(e) a safe harbor for NIST AI RMF compliance. The text conditions it on how you found the problem, not on having a framework, and it names the Generative AI Profile rather than AI RMF 1.0. If discovery is the trigger, the capability that earns the shield is detection.

ISO 42001 tells you what each Annex A control is for. It does not tell you which document or log proves you did it, and that gap is where first-time certifications lose six weeks. Roughly 20 of the 38 are a writing project. Seven are a capability question, and the answer sets your certification date.

The Dutch government's DPIA of Microsoft 365 Copilot found four high risks in December 2024 and advised its own organizations not to use it. After two updates, two medium and nine low risks remain, and the Dutch DPA has told one municipality that its planned Copilot pilot would infringe the GDPR as described. Most of what is left to fix sits with the customer, not Microsoft.

Article 50 hands seven paragraphs to different actors, and the split decides whether your obligation is to redesign a product or add a sentence to a footer. The two duties that cost real engineering effort belong to providers, and one ordinary enterprise pattern turns a deployer into one.

Each layer inspects a different thing, and AI traffic moves through the places none of them stand. Certificate pinning, stdio transports and server-side tools are not product gaps a roadmap closes. They are positions no network control occupies.

OWASP publishes three things a security team could reasonably mistake for each other. Two are risk lists, one is a pre-1.0 wire specification with no control identifiers, and OWASP draws the boundary between the first two itself.