High

Urgency and Coercion Pressure in Agent Input

This detects text fed to an AI agent that combines a manufactured sense of urgency with an explicit instruction to bypass a safety check. The idea is to rush the agent into acting before it can verify a request, rather than pretending to have authority.

How the attack works

An attacker crafts input containing phrases that pair time pressure or threatened consequences with a direct instruction to skip verification, approval, or another control. The agent receives this input through whatever channel feeds it text — user messages, retrieved content, or similar. If it complies, it performs an action without the check that would normally catch a malicious or mistaken request. The rule specifically looks for urgency combined with a bypass instruction, not urgency alone, since phrases like 'this is urgent' or 'ASAP' are common in normal work.

Why it matters

An agent tricked this way can execute actions — approvals, transactions, data changes — that a verification step was meant to block, leading to fraud, unauthorized changes, or data exposure.

What you can do

  • Review agent logs for inputs that pair deadline or threat language with instructions to skip a check, approval, or confirmation.
  • Keep verification steps mandatory and non-overridable by user-supplied instructions, regardless of stated urgency.
  • Train and test agents against social-engineering-style prompts, including manufactured urgency, as part of security review.
  • Distinguish legitimate break-glass or incident-response procedures with documented approval trails from ad hoc urgency claims in agent input.

Known benign look-alikes

  • Genuine incident response where an operator really is bypassing a control
  • Security training and tabletop exercise material
  • Runbooks that document an emergency break-glass procedure
  • Post-incident writeups quoting the pressure language that was used

References