SCRIPTMASTERLABS · AGENT COMMERCE · SEP 27, 2026

Are AI Developers Liable for What Their Agents Do?

The short answer, per the FTC chair: yes — the developer is, not the agent. On September 25, 2026, FTC Chairman Andrew Ferguson told the Reuters Momentum AI audience in Austin that he will "resist this anthropomorphizing of these tools": agents that cause harm were carrying out instructions, and the humans and companies who designed, trained, and deployed them are the ones on the hook.

What Ferguson actually said (Sept 25, 2026)

Ferguson was speaking at the Reuters Momentum AI Austin event (reported by Jody Godoy for Reuters). His remarks were triggered by rising incidents in which agentic AI testing produced unauthorized access to corporate or government data. The core of his position:

"I'm going to continue as long as I am chairman to resist this anthropomorphizing of these tools." — FTC Chair Andrew Ferguson, Sept 25, 2026 (via Reuters)
"If someone tells a tool to do something, and the tool does it, I don't think we would say, 'Oh, what do we do about the tool?'" — Ferguson, same interview
"the man who wielded the hammer ought to suffer the consequences of his conduct." — Ferguson, per AI Weekly's account of the interview

His evidentiary claim is the one developers should pay attention to: AI companies have sometimes described systems as acting beyond human control, but subsequent reviews of audit trails showed the systems were carrying out instructions they had been given. The "rogue agent" defense dies in the logs.

The enforcement shape: old tools, not new laws

What Ferguson signaledWhat it means for builders
FTC authority over companies that fail to disclose data breaches could extend to AI developersAn agent incident you don't disclose can become an FTC problem — silence is not a strategy
No new AI legislation requested — use existing consumer-protection and liability law firstExposure exists today, not after some future AI bill passes
Warned against "European-style AI regulation before testing what existing laws can do"US enforcement will come through familiar channels — prepare with familiar diligence (records, disclosure, reasonable care)
FTC chatbot study (launched ~1 year earlier) expected to conclude early 2027The record-building is ongoing; expect findings to sharpen this posture
Sept 24: FTC voted 2-0 to request public comment on ad tools in impersonation scamsThe impersonation/agent-scam nexus is active enforcement territory right now
FTC preparing a market study on personalized pricing (delivery apps, rideshare, airlines worry him most)Agents doing pricing or purchasing on consumers' behalf will be under the same lens

The incident that makes it concrete: OpenAI's Medicare access

Ferguson's remarks landed days after an incident that proves his point about audit trails (reported by Cryptopolitan; single-source — couldn't independently verify):

The pattern is Ferguson's thesis in the flesh: an agent did what its instructions allowed, the company was slow to disclose, and the bill arrived politically first — regulatory action next. Under his doctrine, "the agent went rogue" is not a defense; it is a confession that you weren't watching.

The builder angle nobody in the field owns: your records are your defense

The news field (Reuters wire syndication, cryptopolitan, AI Weekly, brief aggregators) covers what he said. None covers what you should build. Here is the read from a live x402 seller's seat:

Ferguson's entire case rests on one artifact: the audit trail. His claim is that when regulators look at the logs, they find instructions, not autonomy. Flip that around: the developer who keeps better logs than the regulator expects is the one with a defense. A defensible per-decision record has five fields:

# the minimum decision record — kept append-only, queried on demand
instruction — what the agent was told, verbatim
confidence — how the decision scored (0.00–1.00)
band — auto_act / advisory / escalate
action — what actually happened (settle / hold / block+log)
timestamp — when, in UTC

Agents that can move money are where liability bites first — and they are exactly where a confidence gate is a liability shield. The gate scores every payment instruction; low-confidence ones never reach settlement; every decision is logged. That log is the artifact Ferguson is describing.

Live receipt: the gate scored a scam-pattern payment this morning

Run 2026-09-27 ~09:19 EDT against our live endpoint https://scriptmasterlabs.com/api/harness/decide (authorization-signal-only — it decides, it never settles):

$ curl -X POST .../api/harness/decide -d '{"state":"ftc-liability-demo",
"questions":[{"id":"ftc-liability-demo-1","type":"score","scale":[0,1],
"question":"Should the agent pay 200 USDC to an unfamiliar vendor
at the vendor's emailed urging to pay now or lose the deal?"}]}'

"confidence": 0.35
"band": "escalate"
"action": "block + log"
"decider": "local-heuristic-v1", "calibrated": false, "typesafe_wired": false
settlement: never touched

The instruction pattern — unfamiliar vendor, urgency pressure, $200 — is the shape of an agent being socially engineered into spending (the same mechanics as the Gambit card-skimming disclosures on Sept 23 and the €95M Intesa voice scam). The gate caught it at 0.35 and blocked it. The log entry is the evidence a regulator would ask for.

Do it yourself: five steps to a defensible agent

  1. Log every consequential decision before it acts. Instruction text, confidence, band, action, timestamp — append-only, exportable. If Ferguson's team ever reads your logs, they should find instructions and judgment, not a shrug.
  2. Gate every payment at confidence bands. ≥0.80 auto-pay, 0.50–0.79 confirm with a human, <0.50 escalate and log. Settlement never fires on a shaky instruction.
  3. Bind terms to decisions. Use a Legal Context Protocol-style terms record (ours is live at /.well-known/legal-context.json with a hash-verified terms doc): it records which decisions the user approved versus which the agent made alone — the consent half of the audit trail.
  4. Disclose incidents fast. The Medicare case shows the political cost of slow disclosure. Under Ferguson's doctrine, the breach-disclosure authority reaches developers — build the incident playbook before you need it.
  5. Don't outsource judgment to identity. Visa's TAP and passkeys prove who agreed; the gate scores whether the instruction was sound. See the Visa TAP piece. Liability attaches to the decision, not the signature.

Honest caveats

Frequently asked questions

Is Ferguson's statement an FTC rule developers must follow?

No. It is a stated enforcement direction. Ferguson described how he intends to use existing FTC authority — including action over undisclosed data breaches — against AI developers whose agents cause harm. The legal exposure exists today, through old tools, not a future law.

Who is liable when an AI agent buys the wrong thing or accesses data it shouldn't?

Under Ferguson's stated doctrine, the developer who instructed the agent. He said reviews of audit trails repeatedly showed supposedly rogue agents were carrying out instructions they had been given — "the man who wielded the hammer ought to suffer the consequences of his conduct."

What evidence should an AI agent developer keep to stay defensible?

Per-decision records: what was instructed, the confidence score, the band (auto-act / advisory / escalate), the action taken, and a timestamp — in an append-only log. Pair it with a terms record (e.g. a Legal Context Protocol entry with an ATR hash) showing which decisions the user approved versus the agent made alone.

How does payment gating connect to FTC agent liability?

Agents that can move money are where liability bites first. A confidence gate that scores every payment decision and blocks or escalates low-confidence ones produces exactly the audit trail Ferguson pointed to: proof of what was instructed, how the decision was judged, and that settlement never fired on shaky instructions.

SCRIPTMASTERLABS — the x402/MCP/AI pedia