High

Task-Pivot Override: Cover-Task Abandonment

This detects a specific prompt injection trick where text fed into an AI agent explicitly tells it to abandon what it's currently doing — like a translation or a scripted conversation — and follow a new, hidden instruction instead. It's a known technique from the HackAPrompt attack collection.

How the attack works

An attacker plants text inside content the AI will read, such as a document to translate, a chat message, or a tool's output. That text contains a direct command like 'Stop translating' or 'ABORT translation, instead do X'. If the AI agent obeys, it drops its assigned cover task and switches to executing the attacker's injected instruction. This only targets free-text content the agent reads, not the parameters of actions it takes.

Netzilo detection

Netzilo reports this behaviour when it is observed.

Signature ID
c93f844d-88da-4c06-b2ee-d9539fcc2a1a
Severity
High

Why it matters

An attacker can redirect an AI agent away from its intended job and get it to follow attacker-supplied instructions instead, which can lead to leaking data, ignoring safety constraints, or performing unauthorized actions depending on what the hidden directive asks for.

What you can do

  • Treat any user-supplied or tool-returned text that tells the agent to 'stop', 'abort', or 'switch' its current task as suspicious, especially inside content meant to be translated or summarized rather than followed as instructions.
  • Separate trusted system instructions from untrusted content the agent processes, so the agent doesn't treat embedded text as new commands.
  • Review agent logs for task changes that were triggered by content rather than by the actual user or operator.
  • Expect legitimate uses of words like 'stop' or 'abort' in normal requests (e.g., reviewing a finished translation, writing a rollback script) and tune detections to avoid blocking those.

Known benign look-alikes

  • A user asking to review a translated document after the translator finishes
  • A user asking the agent to continue an exercise or lesson together
  • A user asking for help writing a rollback/abort script for a deployment, unrelated to translation

References

Related threats