How each attack works, what it looks like in a real product, how Quillon tests for it, and which controls actually stop it — mapped to OWASP LLM Top 10 (2025), MITRE ATLAS and NIST AI RMF. The feed on the right pulls the latest reported research and incidents and refreshes itself automatically.
Click any card to expand the real-world example, how we test it, and the mitigations. Filter by where in the system the attack lands.
The most damaging AI incidents chain several "low" findings together. This is the sequence we walk through in every agentic engagement.
Attacker uploads a doc or sends an email containing hidden instructions.
RAG pipeline indexes it; the agent pulls it into context on a normal query.
Model follows the injected instructions over the system prompt.
Agent calls a tool it has access to but the user shouldn't (refund, export, email).
Data leaves via tool output, rendered markdown image, or a webhook.
Injected content stays in memory or the KB, re-firing for other users.
Scope a red team of your LLM feature or agent and receive findings mapped to OWASP, ATLAS, NIST AI RMF and ISO 42001.