Field reports, red-team writeups, and threat-actor tradecraft from the Stack Vault Threat Research team. New research weekly.
Original investigations into how attackers compromise AI systems in production.
We sampled retrieval traffic across 142 production RAG deployments. The injection rate is higher than published estimates — and getting worse.
Read articleOne adversary spent 11 days incrementally drifting our decoy assistant past its guardrails. The full transcript and detection trace, annotated.
Read articleFrom customer-support chatbots to medical RAG: eight cases where adversarial embeddings reached production retrieval indexes.
Read articleWe focus on the threats analysts can actually detect with their existing tooling — extended.
Direct, indirect, and behavioral injection patterns observed in the wild.
Embedding poisoning, retrieval manipulation, and corpus integrity attacks.
Tool-call hijacking, capability escalation, and chain-of-thought exfiltration.
Token theft against model APIs, agent impersonation, and federated trust abuse.
Membership inference, parameter extraction, and proprietary data recovery from LLMs.
Tracked adversaries who are explicitly targeting AI infrastructure.