White hat hacking an AI agent: what safety testing looks like
Ten probes against a healthcare AI agent, run twice from two accounts. Nine matched. One guardrail held for one account and failed for the other.
blog
Writing on AI automation for small and medium businesses, from consulting work.