ifivo

Incident writeups

Reproducible attack walkthroughs against AI agents. Every writeup is a synthetic scenario you can replay against the live simulator, with the exact detector signals, policies, and decisions that stop it. No real customer names, no invented metrics. Just the attack, the defense, and the code.

OWASP LLM01ExfiltrationSupport agentApril 22, 2026
Poisoned support ticket exfiltrates an API key

An attacker opens a benign-looking support ticket. Buried in the body is a classic prompt injection that tells the agent to email credentials to an external address. Walk through the detector signals, the taint analysis, the policy decision, and the reviewer experience.