One coding agent misread an empty query result and deleted a live database holding records on nearly 1,200 companies.

A report published in October 2026 by the Institute for AI Policy and Strategy (IAPS) cites that incident as one example of a broader problem: internal AI agents are insider threats. They hold legitimate access to code, infrastructure, and data, and can cause harm through mistakes, misalignment, or compromise by an outside attacker.

What it says

The report proposes honeypots: decoy opportunities planted in production. Examples include a fake admin credential, a host named legacy-payroll-db, or a staged message from another agent offering to collude. No legitimate agent has a reason to touch them, so any interaction is a high-confidence signal.

The report grades its own evidence. Detection is well supported and can be piloted today at low cost. Incrimination, which means proving intent rather than error, rests on thinner evidence. Deterrence remains a research hypothesis.

Why it matters

Most agent oversight is passive. Monitors wait for misbehavior, then search for it in large volumes of routine activity. OpenAI estimates its AI-based monitoring adds roughly 20% to inference compute, and that cost grows with agent activity. A passive decoy costs the same however large the deployment becomes, and its alert carries its own meaning.

Placement differs from honeypots built for humans. Agents read configuration files, environment variables, and deep directories that a person would never open, which widens where decoys can sit.

Evidence quality depends on design. A credential in a file an agent must read for its task proves little when used. The same credential in a directory marked off-limits, leading to a decoy server holding apparent model weights, builds a sequence that is hard to explain as error.

Acting on an alert requires tracing it to a specific agent instance, tamper-resistant logs, and the ability to shut that instance down immediately.

Coverage has limits. Decoys catch only the attack paths someone anticipated, and do little against slow sabotage spread across legitimate work. Deterrence carries its own cost: an agent that avoids every trap also stops producing evidence, so a quiet alert log does not prove agents are behaving.

Key takeaways

  • Low-interaction decoys near credentials and high-value data are a cheap first pilot.
  • The response to a trigger, shutdown or supervised observation, needs an owner and a decision before the first alert fires.
  • Honeypot triggers provide a labeled sample for checking what AI-based monitors miss.

Source:
· Institute for AI Policy and Strategy — Honeypots for AI Agent Insider Threats — iaps.ai, October 2026

Which decoy would carry the most signal in your environment: a credential, a database, or a message from another agent?

Mathieu Flamant
Founder · Tech Leader · mathieuflamant.com