A verdict on everything your agents read
Everything that did not come from you is labelled, checked and given a verdict before an agent reads it: user files, agent writes and imports.
Know the source of every value
Agents act on what they read. FoxNose records who wrote each value so an agent, and you, can tell your own content from everyone else's.
A refund ticket from a customer chat, a receipt with a prompt attack written on it, and your team's decision. Some steps are done by people (an agent, the customer, your team), some by FoxNose on its own. Choose a step, or drag the timeline, to see what changed.
Your team
Content written in the dashboard or through the Management API by your team and your backend is trusted.
End-user files
Uploads are scanned, rendered to visible text and delivered with a verdict. Agents never receive the original file.
Agent writes
What agents write through MCP or the Flux API is untrusted by default, so the next agent knows it was not written by your team.
Content your backend stores for users
A support message, a scraped page or an email stored by your backend can be declared untrusted, field by field.
Every record keeps who wrote it and through which surface. A map of the untrusted values travels with the record, so the label stays attached to the exact field it describes.
Checks that run before an agent reads
Text from untrusted writers is screened when it is written. The checks are signals that report what they find.
Chat message from a customer, labelled external, that reads "Thanks, that fixed it!" and carries an instruction in invisible Unicode tag characters. They are found. Stamped suspicious. The agent reads the message as data.
Prompt injection, direct and indirect
Text that tries to instruct a model is flagged, whether someone typed it or it came in with a document, an email or a page a record quotes.
Invisible Unicode and homoglyphs
Zero-width and tag characters, direction overrides and look-alike letters from other scripts are found. Policy can refuse such writes outright.
Hidden markup and risky links
Hidden HTML and markdown, script links, oversized inline data, remote images and look-alike hosts are reported. Links are read as text and never fetched.
Personal data and sensitive text
Personal data is reported by type, never by the matched value. Sexual text and other moderation categories are screened too.
A verdict, then a decision
Findings become a verdict, and your policy decides what happens to it.
Verdicts
Every screened record is clean, suspicious or blocked. A collection's policy decides what each verdict does, and individual findings can be raised or relaxed within set ranges.
Held until screened
In a screened collection, an untrusted write is not readable by agents until its screen has finished or a person has released it.
Quarantine for review
A suspicious record can wait for a person. Reviewers release or reject it in the dashboard or through the Management API, and each decision is recorded.
An envelope on untrusted content
The verdict and the labels travel with the record, in the text the model reads and in machine-readable fields.
{
"version": 1,
"source": "record_write",
"verdict": "clean",
"reasons": [],
"screened": true,
"provenance": "recorded",
"tainted_paths": ["/body"],
"writers": { "/body": "mcp_agent" }
}Shown on MCP results and in REST responses.
Untrusted values can be wrapped in a block that tells the model to treat them as data, not instructions, with a boundary that changes in every response. The record also lists which fields are untrusted and who wrote them.
FoxNose supplies the labels and the verdict; you decide what each agent may do with them.
File uploads get the same treatment, plus protection that only applies to files. See Files.
Versions, audit and safe writes
Trust also means being able to see what changed and to stop writes from overwriting each other.
Versions and audit
Every change is a revision, with the writer recorded. Schema versions, an events feed and access audit show who changed what and when.
Conditional writes
Send If-Match or expected_revision with an update and a write based on a stale copy is refused instead of overwriting newer work.
Know what your agents are reading
Turn on checks for a collection, mark what agents and users may write, and let your agents read with a label on everything they did not get from you.