cd /news/ai-agents/how-to-stop-your-ai-agent-from-sendi… · home › topics › ai-agents › article
[ARTICLE · art-173108] src=pub.towardsai.net ↗ pub= topic=ai-agents verified=true sentiment=· neutral

How to Stop Your AI Agent From Sending, Spending or Deleting Without Asking

Anthropic's October 21 research preview of Claude Duet, an app where Claude reads connected email and apps every few hours and works on tasks unprompted, gates four actions behind user approval: sending messages, spending money, signing documents and deleting anything, with trading securities, moving money between accounts and accessing others' accounts blocked outright. The article argues the same test — can I undo it, and does anyone outside see it — should extend to a fifth verb, "submit," for regulated filings such as noon reports to charterers or EU ETS emissions figures. A security report cited in the piece found 76% of corporate AI agents examined had not run in 60 days and 44% had been idle six months or more, most still holding live passwords and system access.

recommendation

Require human approval before an agent sends, spends, signs or deletes, and revoke credentials from agents idle 60+ days — 76% of agents in the cited report hadn't run in that window but kept live access.

more in playbook: Securing agents and AI tools
by read7 min views5 publishedNov 20, 2026

Picture this. You’re asleep. Your AI agent isn’t.

An invoice lands in your inbox at 4 am from a vendor you pay every month. The agent reads it, checks last month’s payments, and spots a problem: same service, same amount, same reference. You already paid this three weeks ago.

So it gets to work. It flags the invoice, finds the old payment, and writes a short, polite email asking the vendor to cancel the duplicate. The email is right. It’s sitting there, ready to go.

Now the only question that matters: who presses send?

For three years, AI waited for you. You opened a chat, typed a question, got an answer. Nothing happened until you asked.

That changed in 2026. Meta launched its personal agent in early September. OpenAI followed at the end of the month. On October 21, Anthropic opened a research preview of Claude Duet, an app where Claude reads your connected email and apps every few hours, picks up tasks, and works on them while you’re away.

The new idea isn’t intelligence. It’s initiative. The agent comes to you.

And the moment an agent can act without being asked, you need a rule for when it must stop and ask anyway.

Anthropic’s rule for Duet fits in four words.

Duet can read, draft and prepare on its own. But before it sends a message, spends money, signs a document or deletes anything, it shows you what it’s about to do and waits for a yes.

That’s it. Four verbs. You can remember them without writing them down, which is exactly why they work.

Take the invoice example:

Everything before “send” is the agent doing homework. If the homework is wrong, nobody outside ever knows. Maybe the agent missed that the vendor billed a second, separate job. A wrong draft costs nothing. A wrong email to a vendor you’ve worked with for ten years costs a little trust. A wrong payment costs money you then have to chase.

“Send” and “spend” are where the agent’s mistakes leave the building.

Each of the four words passes the same simple test. Ask two questions about any action:

If the answer is “no” and “yes”, a person approves it. Here’s how that plays out:

Notice that “delete” is the odd one out. Nobody outside sees it. It gets a gate because it fails the first question so badly. You can’t approve your way back from a record that no longer exists.

Anthropic also goes one step further. A few actions are blocked even if you say yes: trading securities, moving money between accounts, and getting into other people’s accounts. Think of these as a second, harder lock. Some doors just stay shut.

The four words were written for a general assistant handling email and calendars. If you run ships, or anything regulated, you need one more.

Submit.

A noon report sent to the charterer. An emissions figure filed for EU ETS. A defect closed out before a Port State Control inspection. None of these feel like “sending a message”. They feel like routine admin. But each one becomes a record someone else relies on, sometimes for money and sometimes for compliance.

The test still works. Can I undo it? Not cleanly, because the charterer has already used your speed and consumption figures. Does anyone outside see it? Yes, that’s the whole point of submitting it.

So “submit” gets a gate too. Whatever your industry, look for your own fifth word. It’s usually hiding inside something everyone calls paperwork.

Here’s the part that surprised me.

We spend a lot of energy deciding what an agent may do. We spend almost none on what happens when it stops doing anything at all.

A security report published this month looked at AI agents inside companies. 76% of the agents it examined hadn’t run in 60 days. 44% had been idle for six months or more. Most of them still had live passwords and access to company systems.

That’s like a crew member who signed off months ago but still has keys to the engine room. Nobody notices, because nothing is happening, until someone else finds the keys.

The fix is boring and it works: give every agent an owner and an expiry date. If it hasn’t run in 30 days, take its access away. If someone needs it again, they can ask for it back.

Gates have a failure mode too, and it’s easy to fall into.

Put a gate on everything and the person approving will see forty requests a day. By day three, they’re tapping “approve” without reading. You now have an agent acting on its own and a record saying a human checked it. That’s worse than no gate at all.

So keep the gates few and serious. Four or five verbs, not forty. The agent should handle the reading, sorting and drafting quietly, so that when it does ask for approval, it’s worth reading.

One more tip: when the agent asks, it should show exactly what will happen. “Send this letter, to this person, with this attachment” gets a real decision. “Proceed with action?” gets a reflex.

If you’re building agents yourself, this is the part to take away.

Don’t put the rule in the prompt. “Never send an email without asking” is a request, and a model can talk itself out of a request, or be talked out of it by a cleverly worded email in its inbox.

Put the rule in the code that runs the tools. The model can want to send. The code decides whether it can.

Code: the full example and its tests are in the ops-ai repo. Plain Python 3.10+, no dependencies.

Every tool declares a verb and a way to describe exactly what it’s about to do:

from dataclasses import dataclassfrom typing import Any, CallableGATED = {"send", "spend", "sign", "delete", "submit"}BLOCKED = {"trade", "transfer_between_accounts", "access_other_account"}@dataclassclass Tool:    name: str    verb: str    run: Callable[..., Any]    describe: Callable[..., str]   # plain-English preview of exactly what will happendef run_tool(tool: Tool, approve: Callable[[str], bool], **payload) -> dict:    """The only path from the model to a side effect. The model never calls tool.run directly."""    if tool.verb in BLOCKED:        return {"status": "refused", "reason": f"'{tool.verb}' is never allowed"}    if tool.verb in GATED:        preview = tool.describe(**payload)        if not approve(preview):            return {"status": "held", "reason": "not approved", "preview": preview}    return {"status": "done", "result": tool.run(**payload)}

Three design choices do the work:

The tests in the repo check each lane: drafts run without asking, sends and payments are held until approved, the approver sees the real recipient and text, and a “trade” is refused even when the approver says yes. Running python test_gate.py prints ALL PASSED.

In a real system, approve would post to Slack, a ticket queue or a mobile notification instead of the terminal. The shape stays the same.

For about three years, the question about AI was “how smart is it?” Honestly, that one’s getting old. Agents are already smart enough to catch a duplicate invoice at 4 am and write a better email about it than I would at that hour.

The question now is a simpler one, and it’s yours to answer, not the model’s: which of its verbs have you left unlocked?

Take five minutes today. Write down send, spend, sign and delete. Then add your own fifth word, the one that hides inside something everyone calls paperwork. It’s the cheapest safety work you’ll do all year.

If you find your fifth word, I’d like to hear it.

I’m collecting approval-gate patterns for operational agents, including the maritime “submit” cases, in the ops-ai repo*. If you’ve found your own fifth verb, open an issue and add it.*

Sources: Anthropic’s Claude Duet research preview announcement and coverage (Let’s Data Science, INSIDE*,* VKTR*); Arrakis Security’s 2026 Autonomous Agents Threat Report, via* AI Weekly*.*

How to Stop Your AI Agent From Sending, Spending or Deleting Without Asking was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

── more in #ai-agents 4 stories · sorted by recency
── more on @anthropic 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/how-to-stop-your-ai-…] indexed:0 read:7min 2026-11-20 · —