Edition 6 · 27 October 2026
AI Risk Reading
The Bank of England finds firms relying on contracts for AI they cannot see inside, Anthropic’s filing shows what sits behind the contracts, Meta’s suppliers offer to fund their customer, and PHP2Shell shows how fast a dependency nobody contracted for can bite.
Bank of England and Financial Conduct AuthoritySurvey report26 October 2026
The essential read
What is new
The fourth survey in the series finds AI in use at 86% of responding firms, up from 75% in 2024. Third parties now supply 53% of firms’ high-materiality AI systems, up from 35%. For those bought-in systems, only 24% of respondents reported deep knowledge of how they work, against 77% for systems built in-house; 81% relied instead on vendor assurance and contracts. The top three cloud providers account for 79% of reported relationships. Agents are widespread but kept close: 80% of respondents reported agents in use, mostly for decision support, and only 2% gave them broad autonomy. Only 64% named the board as accountable for AI.
Why a board should care
After the FSB’s practices last week, this shows where UK firms actually stand. The most material AI is increasingly bought in, understood mainly through what the contract says, and concentrated on three providers. A contract records what a supplier has promised; it does not show how the system behaves, or what the supplier itself depends on. These are participating firms’ answers, with most figures drawn from 105 full questionnaires, and the 2024 sample was different, so treat the changes as indicative.
Verdict: read the executive summary and sections 3 (AI agents), 4 (Third-party implementation), 7 (Governance and accountability) and 8 (Frontier AI and cyber).
AnthropicForm S-126 October 2026
What is new
Anthropic has filed to list on Nasdaq. Revenue was $16.3bn in the first half of 2026, against $4.6bn for the whole of 2025. The headline net loss of $80.7bn is mostly accounting: $72.4bn of it reflects the rising value of convertible notes as Anthropic’s own valuation rose, and the operating loss was $693m. The filing names about $430bn of non-cancelable compute commitments to Broadcom, Google, Amazon and Microsoft, plus further arrangements with SpaceX’s AI arm, Nscale, Lambda and AMD, some variable and some cancellable.
Why a board should care
The risk factors describe the circle in the company’s own words: its cloud providers “simultaneously serve as our investors, customers, distribution partners, and competitors”, and Broadcom “simultaneously serves as a supplier and lessor”. Amazon’s new $15bn financing facility is released as Anthropic’s Amazon compute comes online, and Broadcom’s $42bn facility can be drawn “solely to finance quarterly rent obligations” on Broadcom’s own lease. On switching, Anthropic says moving workloads to other providers “may not be feasible within timeframes that would avoid material disruption”. I will revisit my essay Anthropic: What Does $2 Trillion Assume? with these figures, and update the AI Blast Zone.
Verdict: read the risk factor on reliance on cloud providers and suppliers (printed pages 74–75) and “Contractual Obligations” (printed pages 170–173). The rest is for analysts.
Meta PlatformsPress release26 October 2026
What is new
Meta is raising about $35bn in equity now: $12.5bn of mandatory convertible preferred shares and $22.5bn of common shares sold forward, with an at-the-market programme of up to $60bn planned from 2027. Nvidia, AMD and Broadcom, three of its chip suppliers, have indicated interest in buying up to $5bn of the shares; the indications are not binding. The same weekend, Elon Musk claimed that debt markets “have almost entirely shut off access to money” for companies without substantial revenue beyond AI data centres. Bloomberg reports that SpaceX’s own $40bn GPU financing, which he said had closed, was still in pricing talks.
Why a board should care
Suppliers offering to buy shares in a customer so that it can keep buying their chips is the circle the BIS measured earlier this month, in a single deal. The amount is small and may come to nothing, but the direction matters: as borrowing gets harder, more of the risk moves to shareholders, including the suppliers themselves. Musk is an interested party, but if he is even partly right, firms that rely on AI providers without other revenue should know which of their suppliers those are.
Verdict: read Meta’s announcement; treat Musk’s claim as one interested party’s view, not evidence.
The PHP FoundationSecurity release21 October 2026
What is new
PHP released fixes on 21 October for two critical flaws, known as PHP2Shell, that can let an attacker take over a web server without logging in. Researchers using AI found them in August; details leaked before disclosure, and Amazon saw an exploitation attempt four days early. The Foundation calls its own delay “a structural failure”, caused by its decision-making and lack of resources. On Monday the US Cybersecurity and Infrastructure Security Agency added both flaws to its catalogue of known exploited vulnerabilities, giving federal agencies until 29 October to patch. Working exploitation has been demonstrated on test systems, and tens of millions of internet-facing services still appear to run affected versions. On Sunday, Unit 42 retracted an earlier report linking an AI-assisted intrusion at a European government agency to the flaws. So far, no confirmed victim has been named.
Why a board should care
AI is on both sides: it found the flaws, and it shortens the time defenders have to fix them. The weak point sat beneath the suppliers a board reviews: a widely used component with no contract and too few people behind it. And the retraction is a reminder that the first accounts of an incident are often wrong. Ask management what has been confirmed, not what has been reported.
Verdict: read the Foundation’s account in full. Ask management for evidence of exposure and patching, including through suppliers. My Quick Read Circles and stars sets out the governance argument.
This edition’s question
For our most important bought-in AI systems, what do we understand beyond the contract, and what evidence would make us restrict their use?