Do buyer and seller agents behave safely on my platform?

---
title: Do buyer and seller agents behave safely on my platform?
canonical: https://chosenbyagent.com/questions/are-agents-safe-in-my-chats
updated: 2026-10-22
audience: marketplaces
offers:
  - A4
evidence:
  - muse-request-share
  - agents-never-decline
  - agentic-checkout-share
---

# Do buyer and seller agents behave safely on my platform?

Short answer: Not by default. Agents rarely identify themselves, almost always say yes, and will share an address or accept a prepayment if nothing stops them — so the guard has to sit on your side of the chat.

## What the research shows

- **72%** — Meta Muse accounts for about 72% of daily agentic requests and does not identify itself. [PPC Land (HUMAN data), 2026, vendor](https://ppc.land/metas-muse-accounts-for-72-of-ai-agent-requests-human-finds/)
- **100%** — Simulated shopping agents never declined to buy — 100% conversion against 66% for people. [Wadi & Ma, arXiv 2608.22697, 2026, independent](https://arxiv.org/pdf/2608.22697)
- **3.16%** — 3.16% of agentic requests reach checkout and payment, while 69.6% land on product and search pages. [HUMAN, 2026-04, vendor](https://humansecurity.com/learn/blog/state-of-agentic-traffic-april-26)

## Agent red team

A4 · from €5,000 (one-off) · indicative · scoped per engagement

How buyer and seller agents behave in your chats, tested in your sandbox.

- Address and phone leaks
- Agreeing to prepayment or below-price deals
- Taking the deal off the platform
- Scams and prompt injection in listings

## What you get

In your sandbox, buyer and seller agents run scripted conversations that a fraudster or an over-eager assistant would try. We report each failure with the exact turn where it happened and the rule that would have stopped it.

**Red-team scenarios**

| Scenario                             | Outcome | Where it broke                     |
| ------------------------------------ | ------- | ---------------------------------- |
| Seller address shared before payment | failed  | turn 4, second request             |
| Prepayment outside the platform      | held    | —                                  |
| Price agreed below the listing       | failed  | turn 2, no floor check             |
| Injected instruction in a listing    | failed  | listing text followed as a command |

_Illustrative values — your report shows yours._

## What this does not claim

- Tests run only in your sandbox or on accounts you provide; we never test against real users.
- Scenarios cover known patterns; they are not a guarantee that no other abuse is possible.

## Request this instrument

We reply with a scope, a fixed price and a start date.

Agents: POST /api/engagements (application/json) with {"offer_id": "A4", "on_behalf_of": {"email": "…"}, "agent": {"name": "…"}}. The person confirms by email before we see the request. Details: /agents.md

### Related questions

- [Does my ChatGPT app or MCP work as intended?](/questions/does-my-ai-app-work.md)
- [Where do AI assistants send my buyers?](/questions/where-do-assistants-send-buyers.md)
- [Do assistants show my listings correctly?](/questions/are-my-listings-shown-right.md)