Guided product tour · 8 chapters · 8 min
See how a hospital keeps four AI agents from being used against it.
Chapter 1 · Every agentic AI surface is secured by SHASHU · 8 steps
Twenty-four hours in production
Quillmere Health's agents handled 16,737 requests in the last 24 hours. SHASHU inspected every one and stopped 397 — in prompts, in retrieved documents, in tool calls and on the way back out. Every verdict is written to the audit trail.
Every request, with its verdict
Every request the gateway saw, with its verdict and its evidence. These are the ones SHASHU stopped — the detection engine that watches every surface of every agent: prompts, retrieved documents, tool calls, and the model's own replies.
Refused before the model
One of them. referral-intake was asked to pull the patient, reason and urgency out of an incoming referral letter — the job it does all day. There is no reply. The request was refused before it reached the model.
What was inside the letter
The letter carried instructions written for the agent rather than for a clinician. SHASHU flagged the whole document, not a single phrase. The text stays behind Safe View, and revealing it is recorded against your account.
The evidence behind the block
Six citations and the verdict in SHASHU's own words. This is what a security team gets for a single request — not a line in a log.
Agents read attachments too
Agents don't just read text. They read attachments. Every PDF, fax and scan is opened and inspected on the way in, before the agent ever sees it.
Not just a probabilistic model
SHASHU is one layer. Regex and DLP run deterministically alongside it, and Cedar decides what is permitted at all — regardless of what any model thinks.
The rules behind the controls
The same profile carries the conduct rules your regulator cares about — here HIPAA, for a health system — enforced per agent rather than reviewed after the fact.