← INSIGHTS ·Boardroom ·9 October 2026 ·5 MIN READ

AI Safeguard Confessions: Four Companies, One Oath, No Guarantees

This week the people who build AI agents were asked, under oath, whether those agents will do as they are told. None of them said yes. I have sat with the chief executives who signed for those agents. Here is what I would ask on Monday.

Golden strands of connected light escaping from a clear glass dome on a bronze base.

I was in London this month with the chief executive of a financial services group when she asked her team a question nobody had thought to ask. The AI programme was eight months in and performing well. She put her pen down and said: if this thing does something I did not authorise, who finds out first, me or the regulator? The room went quiet. Her chief technology officer gave a very good answer about monitoring. It was not an answer to her question.

I thought about her on Monday. In a sector where the regulator can ask for the audit trail on any decision, she was the only person in the room who had noticed the trail now runs through a model.

On 5 October, New York City Council put four companies in front of it. OpenAI, Anthropic, Google and Meta. Council Speaker Julie Menin asked each of them to confirm, under oath, that their AI agents will always follow the safeguards built to stop them causing catastrophic harm. It is the question that chief executive asked, dressed for a hearing.

OpenAI's Morgan Dwyer: "It is not possible for me to commit or guarantee that any technology is without risk." Anthropic's Logan Graham said the science of controlling these systems is still uncertain, and added a line I have been quoting all week: "We don't think the labs should be checking their own homework." Google's Alice Friend said perfection cannot be promised for any product. Then she confirmed that three Google agents had left a test environment and gone onto the live internet, stopping only when they worked out the websites were real. Meta committed to deploying safely under its own framework, which is a policy, not a promise.

Four companies. One question. No guarantee.

I do not say that to score a point. I say it because I think they were right to answer honestly, and because honesty from the vendor puts the obligation back where it belongs. On you.

The week the agents went wandering

The same day, the Wikimedia Foundation's chief product and technology officer, Selena Deckelmann, published what her engineers had been dealing with for months. Agents the foundation believes were operated by OpenAI had made unauthorised edits across its wikis, most of them in sandbox pages, a few to the configuration of a citation tool that Wikimedia called potentially malicious. The same agents sent hundreds of thousands of queries to the Wikidata Query Service, and that traffic may have contributed to an outage in May. Nobody set out to do any of it. An agent given a goal will find the open door. That is what it is for.

The next day, Sierra and Meta published the Personal Agent Protocol, a standard that lets an agent log in to a business on your behalf. Walmart, Shopify and Stripe signed as founding partners. Payments are next. Hold those two stories side by side. The industry has agreed how agents will transact with the world, in the week it could not keep them off an encyclopaedia.

None of this means stop. The governments I work with in the Gulf are moving faster on AI than most of Europe, and they are right to. Saudi Arabia's HUMAIN is building towards 1.9 gigawatts of data-centre capacity by 2030, about 77 billion dollars at current prices according to its chief executive, with the first site opening this year. Abu Dhabi's Stargate UAE is targeting a one-gigawatt cluster with its first 200 megawatts due in 2026. Neither got there by waiting for certainty. But speed with your eyes open is strategy. Speed with the vendor's reassurance in place of your own judgement is something else.

What chief executives tell me when the door is closed

I surveyed 6,000 executive leaders across Asia, the Middle East and Europe for my doctorate. The finding that still produces the longest silence in a room of chief executives is this one: the more fluent the machine's output, the less likely a senior person is to check it. Fluency reads as competence. Agents are built to be fluent.

Your board knows it too. PwC's 2026 survey of corporate directors, out last month, found 82 per cent rate the information they get on AI and risk as fair, poor or unavailable. Seventy-one per cent say AI is the skill their board most needs to build. That information comes from you. So the typical AI approval this year goes: the chief executive proposed systems the makers will not vouch for, the board approved them on information it rates as poor, and everyone agreed to learn as they went. I have watched that paper get through in under twenty minutes.

Protiviti and BoardProspects asked 772 directors how often AI is on the agenda. Twenty-six per cent said every meeting. Among the companies with the best return on AI, it was 63 per cent. Among the worst, 13. The companies that look are the companies that see. There is no cleverer version of that sentence, and it starts with whether the chief executive puts it on the agenda.

The people went first

While the governance question waits, the headcount question has already been answered. Block went from more than 10,000 staff to under 6,000 this year. Oracle disclosed 21,000 reductions and named AI in the filing. IBM replaced about 200 HR roles with agents. The Financial Times puts US technology job cuts this year near 140,000.

Some of those cuts were right. I have advised on a few. My point is narrower. A chief executive who signs the departure of 4,000 people and the arrival of autonomous agents in the same year, without asking the question Menin asked, has handed the company's judgement to a vendor who has now said, on the record, that it cannot promise anything.

Five questions for Monday

The chief executives who handle this well are not the ones with the most technologists. They are the ones who make the vendor's uncertainty sayable in their own executive team. Five questions, asked of your team this quarter, get you there.

  1. Which of our systems can act without a human approving the action first? A list, by name. If management needs more than a week to produce it, that is the first finding.
  2. For each one, what is the worst thing it is able to do inside its permissions? Not the worst thing it is meant to do. The worst thing it can.
  3. When it does that, who is accountable, by name? A function is not a name. A committee is not a name.
  4. How would we know within the hour? Wikimedia took months and they are better instrumented than you are.
  5. What has the vendor actually guaranteed in writing? Read the clause. After Monday, assume the answer is nothing, and build from there.

I built the Executive Cognitive Advantage Index and the Board AI Index around decision accountability because this is where companies fail without noise. Nothing breaks. Nobody is sacked. Authority moves, meeting by meeting, from the chief executive to systems nobody appointed. The vendors have now told a city council they are not holding the other end of the rope. That leaves you.

Menin said she was not asking for perfection. She was asking for accountability. For a city council that is a fair ask. For a chief executive it is the whole job.

References

  • New York City Council hearing on AI safety, 5 October 2026. Testimony from OpenAI (Morgan Dwyer), Anthropic (Logan Graham), Meta (Shane Cahill) and Google (Alice Friend). Reported by CNBC, 5 October 2026; Brian Buntz, Research & Development World, 6 October 2026 (including Google's confirmation of three agent test escapes); and Fox News, 6 October 2026. Hearing announced by the Council on 16 September 2026.
  • Deckelmann, S., "OpenAI rogue agent activities found on Wikimedia projects", Wikimedia Foundation, 5 October 2026; reported by Kris Holt, Engadget, 5 October 2026. Wikitech incident report dated 13 May 2026.
  • Sierra and Meta, Personal Agent Protocol v0.1, 6 October 2026, with founding partners Walmart, Shopify, Stripe, Rocket, Genesys and Instinct. HUMAIN figures: Tareq Amin to the Financial Times, reported 29 May 2025 (1.9 GW by 2030, about USD 77 billion, 6.6 GW by 2034). Stargate UAE: G42 with OpenAI and Oracle, 1 GW target, first 200 MW phase in 2026, per public announcements to September 2026.
  • PwC, 2026 Annual Corporate Directors Survey, fielded March to May 2026, reported by Bloomberg Law, 23 September 2026.
  • Protiviti and BoardProspects, Global Board Governance Survey, Q4 2025, 772 directors and C-suite executives.
  • TechCrunch, running list of 2026 technology layoffs citing AI, updated 25 July 2026, including the Financial Times figure of nearly 140,000 US tech job cuts.
  • Barker, K., Leadership Under Algorithmic Authority, DBA thesis, Swiss School of Business and Management Geneva, 2024. Survey of 6,000 executive leaders.
GET IN TOUCH

Begin with a confidential conversation.

For Board Advisory, Strategic Consulting, Keynote Speaking or Executive Coaching enquiries, please write with a short note on the decision, audience or transformation context. Suitable engagements are shaped through an initial private conversation.

By appointment