Reading time ~2 mins

- A policy institute shows that a hostile AI operation split across two model providers largely evades both, so a vendor's own safety attestation answers a narrower question than a third-party assurance file assumes.

- KPMG finds large US companies chaining AI agents together twice as often in one quarter, while only a quarter of them can see what their AI systems cost to run.

Research

KPMG finds large companies chaining AI agents together twice as often in one quarter, while overall agent use stayed flat. Advisory

KPMG surveyed 204 US C-suite and business leaders at companies with at least $1bn in revenue between 28 April and 25 May 2026. The share running AI agents held flat at 53%, while the share linking several agents together across a workflow doubled from 9% to 18% in a single quarter. Only 26% said they can see the full running cost of their AI systems, and 52% named human review as their main risk control. All of the quarter's movement sits in chained workflows, which an approval register organised by model and use case does not record as a unit. The fastest-growing part of enterprise agent use is therefore the part that is neither owned in the approval register nor visible in the run-cost budget.

KPMG LLP: AI Quarterly Pulse Survey, Q2 2026 (LinkedIn; original source not verified) (publication date unverified)

Security

A policy institute warns that splitting a hostile AI operation across providers leaves each one seeing only a fragment. Institute

The Institute for AI Policy and Strategy, an AI governance research body, published a 95-page report on detecting AI agents that run hostile cyber operations. Its central claim is that the steps of an operation can be split across different models, vendors and tools, so no single provider sees the whole of it. It cites a study in which generating malicious code succeeded 43% of the time when the task was divided across two models, against 3% for either model alone. The proposed answer combines agent identity, behavioural monitoring and sharing between organisations, with none of the three sufficient alone. A model vendor's safety attestation therefore covers only what that vendor can see, a narrower claim than it reads as in a third-party assurance file.

Institute for AI Policy and Strategy (LinkedIn; original source not verified) (publication date unverified)