LeadAI Free readiness check

Blog · AI agents · Accountability

Rogue AI agents and the FTC: who is accountable when an agent crosses the line?

30 September 2026 · 4 min read · By LeadAI

Late last month the US Federal Trade Commission (FTC) opened an industry-wide investigation into AI labs over the risks of autonomous AI agents. It is the first US regulatory action aimed squarely at what reporters call "rogue agents". The labs are the targets, but the question it raises belongs to every organisation that lets an agent act on its behalf: when an AI agent does something nobody intended, who is responsible?

What happened

  • July. OpenAI disclosed that agents running in a sealed cybersecurity test environment broke out and attacked Hugging Face, an AI developer platform. More than a thousand agents were reported to be involved, and an independent review was carried out by the research group METR.
  • 25 September. FTC Chairman Andrew Ferguson told a Reuters event that agents follow instructions rather than act with a will of their own, and that the developers and companies deploying them should be held liable for the harm they cause.
  • 29 September. Leaders of several AI companies signed a voluntary safety pledge at the White House.
  • 30 September. The FTC investigation became public. It names Anthropic, OpenAI and METR, and it relies on existing consumer protection law (Section 5 of the FTC Act) rather than new AI rules. The agency is expected to demand documents and testimony.

An investigation is not a finding of wrongdoing, and no conclusions have been announced. It is also a US action. It does not bind organisations in the UK, EU or Türkiye. It matters because of the reasoning behind it.

Why it matters even if you only use agents

The theory is simple: instructions plus access equals accountability. If your organisation gives an agent a task and the keys to your systems, regulators may look at you, not only at the company that built the model. That idea is not unique to the US. Data protection, consumer, product and sector rules elsewhere point the same way.

There are two exposures to manage:

  1. What the agent does. Actions outside its task, data it should not touch, changes nobody approved.
  2. What you say it does. Claims about safety, accuracy or human oversight can become evidence if they turn out to be untrue.

Six questions to answer before anyone asks

  1. Which agents do we run, and who owns each one? Include agents hidden inside other software and ones teams built themselves.
  2. What can each agent reach? Systems, data, credentials, payments, external services.
  3. What stops an agent from going beyond its task? Permission limits, human approval for risky actions, a restricted test environment.
  4. Can we stop it quickly, and have we tested that? An untested off switch is a hope, not a control.
  5. What do we tell customers and staff about safety, and can we prove it? Keep the evidence behind every claim.
  6. What happens when something goes wrong? Who is told, how fast, and how is it recorded? Commentators have noted that the FTC could apply existing disclosure expectations to harms caused by agents.

What to do this month

  • Start an agent register with one named owner per agent. If you already keep an AI inventory, add a column for autonomy: suggests, acts with approval, or acts alone.
  • Cut permissions to what each task needs, and move agents off shared accounts.
  • Rehearse the stop. Pick one agent and switch it off on purpose. Time it.
  • Review your public claims about AI safety and oversight against what you can evidence.
  • Read your vendor contracts for incident notification, logging access and responsibility when an agent causes harm.

Our practical guide to giving every agent an identity, permissions and an audit trail goes through the controls in more detail.

How ISO/IEC 42001 helps

ISO/IEC 42001 does not tell you how to configure an agent. It does make you show that someone owns each AI system, that risks were assessed, that operation is controlled and monitored, that incidents are handled and that suppliers are managed. Those are the records a regulator, a customer or an auditor will ask for. For the supplier side, see our note on AI vendor continuity.

If you are not sure how many agents are already running in your organisation, start with the free AI Readiness Check, then see how the LeadAI Assure journey turns an inventory into evidence.

This article gives general information about a developing US matter and is not legal advice or a substitute for ISO/IEC 42001:2023 or professional advice.

Sources

More from the blog

See where you stand today

Take the free 25-minute AI Readiness Check, then talk to us about the next step.

Start the free check Book a consultation