Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

Third-party risk content is written from the customer's side. Assess your provider, tier your vendors, understand your concentration. ‍ A technology provider is on the other end of every one of those assessments, and its own incident propagates outward through contract rather than inward through remediation. The instinct is that the contracts therefore determine the loss. They determine the smaller half of it. ‍

A cyber loss model for a research organization counts subject records and applies a per-record cost. Personal health information, a notification exercise, a regulatory penalty. ‍ The mechanism that matters in a trial is integrity rather than confidentiality, and it produces a loss that occurs even where nothing was altered. What gets destroyed is the ability to demonstrate that nothing was. ‍

The assumption about AI systems affecting children is that the consent structure carries the difficulty. The subject cannot consent, so a parent consents instead, and the governance problem is collecting and tracking that permission. ‍ The assumption is backwards. Consent is usually the wrong lawful basis for these deployments, so the parental consent machinery is not required at all. What differs is something else entirely. ‍

A tool gateway reads every call that crosses it and enforces policy on each one. It is blind to whatever never traverses it, which is established and not the interesting part. ‍ The useful number is what fraction of an agent's total action surface the gateway covers. Blindness of unknown size and blindness of known size are different problems, and only the second lets you state what a gateway-based claim is worth. ‍

The concern about fine-tuning is that it quietly converts a deployer into a provider, pulling in conformity assessment, technical documentation and a quality management system nobody budgeted for. ‍ The concern is misdirected. Fine-tuning is among the least likely routes to reclassification, and the route almost nobody worries about requires no training compute at all. ‍

A public body running a high-risk AI system owes a fundamental rights impact assessment under Article 27 before first use, with the results notified to a market surveillance authority. The obligation is real and it is not yet in force. ‍ Regulation (EU) 2026/1744, in force since July 2026, deferred the section of the Act containing Article 27 to December 2027 for standalone high-risk systems and August 2028 for those embedded in regulated products.

A recruitment system faces two obligations at once. AI rules classify it as high-risk and require the governance that follows. Employment law has required attention to discriminatory effect for decades, independent of any technology. ‍ They are not two versions of the same requirement.

Saying a control reduces exposure is easy and almost always true. Saying it reduces exposure by a specific amount is a different claim, and the machinery for producing one is well established. Set a baseline from frequency and magnitude ranges, simulate, re-estimate the ranges with the control in place, simulate again, and report the difference. ‍ The method is sound. Applied to AI controls it runs into two problems, one about which term the control touches and one about what the estimate rests on.

The OWASP GenAI Security Project unveiled an Agent Control Standard in early September, donated to the project and aimed at runtime enforcement for agentic systems. It sets out that agents should be inspectable, traceable and instrumentable, with declarative hooks and policy enforcement across frameworks. ‍ It answers which controls belong around an agent.

That human review becomes a bottleneck as agents scale is now widely observed. Five or ten agents working in parallel produce more decisions than one reviewer can evaluate, and under queue pressure the review degrades into approval without examination. ‍ The usual response is to move up a level, reviewing intents and boundaries rather than individual outputs.