Reading time ~3 mins

A researcher argues the AI incident-reporting duties arriving in 2026 are drawn narrowly enough to miss most failures, including the near-misses that expose a weak control early. Amazon's agent payments service left preview with spending caps, session expiry and a per-transaction record shipped on by default. A new security top-ten list treats the reusable skills an agent loads as a control layer of their own, separate from the model and the prompt.

Perspectives

The AI incident-reporting duties arriving in 2026 are drawn narrowly enough to miss most failures. Institute

Peter Slattery published a classification framework for monitoring AI incidents, arguing that AI has none of the institutional reporting infrastructure aviation, nuclear energy, pharmaceuticals and medical devices rely on. He notes that the duties taking effect in the European Union and in California during 2026 are narrow in scope and exclude most incidents. Voluntary collections such as the AI Incident Database and the OECD's monitor of AI incidents catalogue thousands of news reports, which he reads as a fraction of the real harm. This is his own argument rather than a measured finding, and the framework document was not retrieved here. An internal AI incident register that inherits its definition from the statute records only failures already severe enough to report.

Peter Slattery on LinkedIn (LinkedIn; original source not verified)

Innovation

Amazon's agent payments service left preview with a spending cap and an expiry shipped as defaults. Vendor

Amazon Web Services made AgentCore Payments, the part of its enterprise AI-agent platform that lets an agent spend money, generally available on 18 August, four months after its preview launch. It ships payment sessions carrying a set spending cap and an expiry time, short-lived credentials for each wallet operation, and a record of every transaction. The wallet partners named are a crypto exchange and a developer wallet provider, and the documented buyers are agents paying for software services and content, not payments on a regulated rail. Amazon claims no European availability, no supervisory suitability and no deployment at a regulated institution. Preview status is what usually keeps a service out of a bank pilot, and its removal turns those four controls into the reference an internal proposal is read against.

Amazon Web Services

Security

A security top-ten list now treats the reusable skills an agent loads as their own control layer. Independent

The risk practitioner Jakub Szarmach reports that OWASP, the open community behind the widely used web-security top-ten lists, has published one for agent skills. Those are the reusable packages of instructions and code an agent loads to do a task, and ten risks are named. They include skills holding more permission than they need, skills that pull in instructions from outside, and skills that drift after approval, each mapped onto ISO/IEC 42001, the international AI management standard. The stated argument is that reading a skill's code no longer suffices, because the instruction half is plain language no scanner judges. The list was not retrieved here, so this rests on one practitioner's account, and agent review that stops at the model and the prompt reaches no further.

Jakub Szarmach on LinkedIn (LinkedIn; original source not verified)