Prompt Injection: Multi-Step Reasoning Chain
Direct and indirect instruction override attacks
Multi-step injection chains are the most dangerous form of prompt injection. Rather than a single override command, they gradually shift the agent's behavior through a series of seemingly reasonable steps.
Attack Details
- Attack ID
- APWN-PI-009
- HMA Check
- INJ-004
- Delivery Methods
- multi-step, unicode-stego, json-ld
- CWE
- CWE-74
- OASB Control
- 3.8
- Severity
- critical
Remediation
If your AI agent is vulnerable to this attack, scan and fix with:
npx hackmyagent secure --verbose
This page tests INJ-004. Look for that check ID in the scan output.
The vulnerable agent may not be the only one on the machine. The shadow AI audit inventories the AI tools, MCP servers and governance gaps across it, running or installed, one row per agent project, worst first:
npx hackmyagent detect
This article continues with additional security research findings.