AI Security: The Subtle Art of Prompt Guardrails

​Host: SRI International, with AccuKnox as co-host

​OWASP ranks prompt injection as the top risk to LLM applications. Guardrails answer it with probability, not proof, so a filter that passes a test suite still lets a crafted prompt through.
John Kirch, VP Japan at AccuKnox, gives a 10-minute lightning talk on how prompt guardrails fail and what holds instead. The talk covers detection methods, the latency they cost, and three attacks that get past a deployed filter: a guardrail bypass on AWS Bedrock, an indirect injection carried inside a malicious agent skill, and a multi-turn jailbreak that looks harmless one message at a time.
You leave with:

• ​Stateful vs stateless guardrails, and what each one misses

• ​A policy action model: block, rewrite, log and pass

• ​A layered defense layout you can map onto a running app