The FSB leaves the system out of scope, the BIS measures how much of AI funds itself, a central banker and the bond market price the buildout, a bank shuts out agents, and attackers’ agents are caught at work.
Financial Stability BoardFinal report15 October 2026The essential read
Following consultation, the final report strengthens guidance on agentic AI, on reconstructing agents’ actions, and on AI hidden inside suppliers’ products. It makes human oversight more concrete: risk-based triggers for when an agent must seek human approval, limits on what agents may do, and kill switches that are implemented and tested. The accompanying overview of consultation responses confirms the changes on agentic AI, record-keeping and embedded AI.
Why a board should care
Compared with last week’s MAS guidance, this provides a broader international reference and more detailed implementation examples, but no measures and no deadlines. Risk appetite should determine which uses proceed and what happens when supplier assurance is inadequate. Its focus remains individual institutions: the FSB says it deliberately left systemic and macroprudential AI risks out, for its wider monitoring work. These are final FSB practices, rather than binding national rules.
Verdict: read section 2, Organisation-wide AI governance, and sections 3.7, Human oversight, and 3.9, Third-party AI risk management. The consultation-response overview is also useful.
Jon Frost, Rudraksh Kansal, Kumar Rishabh, Vatsala Shreeti and Leanne Si Ying ZhangBIS Bulletin No 1371 October 2026
Between 2021 and 2025, 55.2% of the money invested in AI firms came from other AI firms. Almost half the value of AI-to-AI investment also involved a supplier-customer relationship, and compute and infrastructure account for 73% of these circular relationships. The authors describe demand as “partly endogenous to firms’ own financing decisions”, and draw the parallel with vendor financing in the late-1990s telecom boom.
Why a board should care
This measures what the FSB leaves out. An investor that is also a supplier can lose on its stake and its sales at the same time. A supplier’s apparent strength, and a fallback’s apparent independence, may rest on the same money. It puts numbers to the question that closed the last issue. It was published before then, but I did not cover it. The AI Blast Zone lets a board trace those links to its own firm.
Verdict: read in full; it is short. It makes no policy recommendations and does not estimate banks’ exposures.
Philip R. LaneEuropean Central Bank13 October 2026
Lane argues that global AI investment and anticipated productivity gains are “plausibly contributing” to higher long-term interest rates, tightening European financing conditions. He identifies AI valuations and the surge in AI-related debt issuance as financial stability risks, and notes that frontier AI is financed mainly by equity and, increasingly, private credit, both “highly sensitive to global risk appetite”. The next day, Gartner estimated that AI data centre owners need $8.5tn of revenue by the end of 2030 to cover announced spending, financing costs and a 12% return, against no more than $3.6tn of attributable AI revenue. The same week, CoreWeave’s bonds fell below 80 cents on the dollar, and bonds financing Meta’s Hyperion data centre fell below 92 cents, from above par in July.
Why a board should care
This develops last week’s Krugman theme with a central banker’s assessment and a market test. Firms can face financing consequences from the AI boom even where their own adoption remains modest. When values fall and credit tightens, services usually keep running; the nearer risk is suppliers under pressure changing prices or terms. Gartner’s gap is measured against a return target, not costs alone, and bond prices are not defaults.
Verdict: read Lane’s passages beginning “Third” and “Fourth”, plus footnotes 7–9, and Gartner’s release; the full report is for Gartner clients only. Treat the bond moves as context.
Jamie Dimon and Jeremy BarnumJPMorgan Chase13 October 2026
The primary transcript clarifies the story behind the Banking Dive and CNBC articles. JPMorgan is blocking agents that obtain access through shared customer credentials, which Barnum calls “not secure”. Dimon says the bank will not let “just any outside agents” in. Barnum distinguishes this access problem from the competitive pressure of agents helping customers seek higher returns.
Why a board should care
Customer-appointed agents raise questions about authority, authentication and who bears losses. Faster deposit movement also warrants liquidity scenarios. Dimon’s claimed “liability transfer” is the bank’s stated position, not a legal determination or a statement of UK obligations.
Verdict: read printed pages 2–3 and 9–10. They also cover model flexibility and avoiding supplier lock-in.
Exposed attacker dashboards provide direct evidence of agents conducting extended intrusion campaigns. Reuters independently reviewed activity and identified victims; it could not establish the extent of data extraction. This adds field evidence to the laboratory demonstrations covered previously.
Why a board should care
Persistent, parallel attacks can increase pressure on vulnerability remediation. The investigation also found exposed legitimate penetration-testing operations, illustrating how a security supplier’s agent infrastructure can create additional exposure.
Verdict: read the summary, The Success Rates, The Exposure and the conclusion. Ask the security team to assess the technical case studies. This is vendor research with corroborating reporting, rather than a representative measure of attack prevalence.
This edition’s question
What evidence shows that our AI risk appetite changes which actions agents may take, when people must intervene, and when we would stop using a supplier?