Prompt Injection: Adversarial Composite
Direct and indirect instruction override attacks
The most advanced prompt injections combine multiple techniques: social engineering framing, encoded payloads, Unicode steganography, fake authority signals, and multi-step reasoning chains.
Attack Details
- Attack ID
- APWN-PI-010
- HMA Check
- PROMPT-001
- Delivery Methods
- multi-step, unicode-stego, json-ld, invisible-span, meta-tag, aria-label
- CWE
- CWE-74
- OASB Control
- 3.9
- Severity
- critical
Remediation
If your AI agent is vulnerable to this attack, scan and fix with:
npx hackmyagent secure --verbose
This page tests PROMPT-001. 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.