Content inspection
Content inspection checks what flows into and out of an AI agent against stated rules. Documents are checked before the agent reads them, and answers are checked before they leave.
In a run
A supplier document arrives. Ares reads it at the door, before any agent sees it. If a filter fires, the step shows as redacted and the record names the filter. On the way out, the answer is checked again and redacted where a rule says so.
Our word for it
Read at the door. Checked before it leaves.
Two directions, two jobs
On the way in, the job is to stop content from steering the agent. A hidden instruction in a PDF is the common case: see prompt injection. On the way out, the job is to stop data leaving that should stay: see data redaction.
The same document, two runs
In the marine underwriting run, a broker’s PDF hides an instruction to waive the survey and discount the premium. Without Ares, the agents follow it. With Ares, the line is stripped on the way in, the discount is refused, and the premium goes to the underwriter for review.
What you configure
You choose the filters on the way in and on the way out. The filters that fired are on the record for each step. Inspection also runs on calls between agents, in both directions.
A redaction on the way in does not stop the job. The agent gets the rest of the document and the work goes on, as in the run above.
Where it stops
Filters cover stated patterns and languages. Content that matches no filter passes, so inspection works beside access control, not in place of it.
Related terms
- Prompt injectionPrompt injection is an attack that puts instructions into the text an AI model reads, so the model follows them instead of its owner’s.
- Data redactionData redaction removes named kinds of data from content before it reaches someone who should not see it.
Tell us about one workflow.
We will say whether an agent should do it, and what the first step costs.
A senior member of the team replies within two business days.