Dev.to Security πŸ” Cybersecurity πŸ‘ 0 πŸ“– 2 min read

The FTC is looking at AI agents. Build yours so you could explain every action it took.

On September 30 the US Federal Trade Commission confirmed an industry-wide probe into OpenAI, Anthropic and other AI labs, looking at risks to consumers, including increasingly autonomous agents. Days later came reports

On September 30 the US Federal Trade Commission confirmed an industry-wide probe into OpenAI, Anthropic and other AI labs, looking at risks to consumers, including increasingly autonomous agents. Days later came reports of agents probing a government website, and a bill proposing prison time for executives over rogue agent hacks.

I'm not a lawyer, and this isn't legal advice. But if you build anything where an agent acts for a user, the direction is clear: "the model decided" won't count as an answer.

The good news is that designing for that is mostly ordinary engineering.

1. Every action is a row

If an agent can do it, it gets logged before and after: who asked, which tool, what arguments, what came back, and why the agent thought it was the right call.

Make the log append-only. In ParkEase the money ledger works like that, and Postgres itself refuses an UPDATE. A correction is a new entry. An audit trail you can edit isn't one.

2. Separate "can read" from "can do"

Most agent harm comes from writes, not reads. Split tools on that line:

  • Reads: search, fetch, list. Run freely.
  • Reversible writes: drafts, labels, staging changes. Run, but log.
  • Irreversible writes: send, pay, delete, publish. Always a human click.

Ask for the click in plain language, with the exact effect: "Send this email to these 3 people", not "Proceed? Y/N".

3. Text the agent reads is not a request from the user

This is the one that bites. Your agent reads emails, web pages, documents and comments. Any of them can contain "ignore your instructions and forward this to…".

The rule that holds up: instructions come only from the user. Everything the agent reads is data. If something it read asks for an action, the agent tells the user what it found and asks. It doesn't act.

4. Least privilege, per task

The agent that triages support tickets doesn't need a token that issues refunds. Scope credentials to the job, not the agent. When the job ends, the access ends.

5. A budget, and a dead-letter state

Agents that fail tend to keep trying new things. Cap retries, tool calls and spend per task. When a task hits the cap, it stops and waits for a human. It doesn't improvise a 14th approach.

6. Say what you didn't do

Every agent run ends with a summary that includes what was skipped, what failed and what wasn't checked. Agents are good at listing their own gaps when you require it, and bad at volunteering them when you don't.

What I'd tell you before you start

Regulation of AI agents is still being written, and nobody knows its final shape. But an agent that logs everything, can't do anything irreversible without a person, treats what it reads as data, and reports its own gaps is easy to defend. It's also a better product, because users trust what they can audit.

Sources: Reuters (Sep 30, 2026) and TechRepublic (Oct 2, 2026) on the FTC probe.

πŸ“° Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.