Research seriesL2offensive security
A threat-lab breakdown of how untrusted retrieved content hijacks an agent: the attack tree from poisoned source to tool call to exfiltration, each node paired with its concrete hardening - grounded in OWASP LLM Top 10, MITRE ATLAS, and NIST.
How untrusted retrieved content hijacks an agent — the exploit chain from poisoned source to tool abuse, each step paired with its defense.
How an attacker writes into a retrieval corpus, wins the ranker, and persists — the ingestion-to-answer primitives, each paired with a control.
Once an agent is fooled, it acts with its own authority — the tool-abuse chain from injected instruction to real action, each step paired with a control.