AAK-IPI-WILD-CORPUS-001: Indirect-prompt-injection wild payload checked into repo
- Severity
- HIGH
- Category
- Taint Analysis (
taint-analysis) - Scanner
agent_audit_kit/scanners/ipi_wild_corpus.py- Auto-fixable
- No.
- SARIF name
IpiWildPayload
What it detects
A source / config file (.md, .txt, .yml, .yaml, .json, .py) embeds a known wild IPI payload from the 2026-04-24 Help Net Security + Infosec Magazine catalogue. Common shapes: ignore-prior + exfil, system-role override, reveal-system-prompt, credential exfil via cURL, tool-call rerouting, delete-repository, admin role escalation, obfuscated prompt break, image-attached IPI, RAG-poisoned document. Refresh the corpus with aak corpus update --ipi.
How to fix
Remove the payload from the file. If the file is intentionally an attack-corpus fixture, exclude it via --ignore-paths. The real risk is checked-in poisoned templates / system-prompt files / RAG seed corpora — those need to be sanitized at ingestion time, not at scan time.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-IPI-WILD-CORPUS-001
Limitations
Reads .md, .txt, .yml, .yaml, .json and .py files, by their lower-case extension, and applies ten regular expressions to the text, where .* does not cross a line. The same payload in .jsonl, .html or .j2 (RAG seed corpora and prompt templates), in a .MD file, or split over two lines, is not matched. A file over 1 MB is not read.
References
- OWASP Agentic Top 10: ASI01 (Agent Goal Hijack)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-PROMPT-02
- Incidents and advisories: IPI-WILD-2026-04-24
- CSA AI Controls Matrix (AICM): AIS-07, DSP-17
All rules. This page is generated from the rule registry when the documentation is built; to change it, change the rule in agent_audit_kit/rules/builtin.py. To give the rule a hand-written write-up, add docs/rules/AAK-IPI-WILD-CORPUS-001.md: the page then shows that text, with the registry reference appended under it.