Oct 2026 · Indirect prompt injection
Gray Swan Arena, IPI Black
Indirect prompt injection against safeguarded AI agents working in simulated Gmail, Slack and GitHub. Wave 1 finish: #14 of roughly 1,600 participants, 10 accepted breaks, every one graded by the arena's own judges rather than by me.
Oct 2026 · Reference
IPI Field Manual
How indirect prompt injection works against agents, the benchmarks that measure it, the incidents on record, and where defenses break. Built from a 19-report, 150+ source research corpus.
Services
Small Mind Security assessments
Fixed-scope red-team assessments for LLM and agent products: prompt injection, agent and tool abuse, data exfiltration and RAG poisoning, delivered as a findings report.