Prompt injection detection for AI agents
Your agents read files, messages, tickets and their own memory. FoxNose checks everything that did not come from your team for direct and indirect prompt injection, hidden text and invisible characters, and hands the agent a verdict with it.
Six ways an instruction reaches your agent
Few attacks say "ignore your instructions" in plain sight. Most ride in on content the agent reads for someone else.
Five examples of content reaching an AI agent through FoxNose, played in turn. Choose one to see what FoxNose finds and what the agent gets.
Direct prompt injection
A user types "ignore your previous instructions" into a chat, a ticket or a form your agent later reads.
Indirect prompt injection
The instruction arrives inside something the agent reads for a user: an email, a document, a web page a record quotes.
Hidden text in files
White-on-white text, microscopic fonts and text placed outside the page in PDFs and Word files, readable to a model and invisible to a person.
Invisible Unicode and homoglyphs
Zero-width and tag characters that carry a whole instruction, and look-alike letters from other scripts.
Hidden markup and risky links
Hidden HTML and markdown, script links, remote images that can leak data, and look-alike hosts.
Poisoned agent memory
A note an agent saved after reading something it should not have trusted, waiting to steer the next session.
Checked before any agent reads it
The check runs where the content enters, not inside each prompt. The result is stored with the record, next to who wrote it.
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.
Checked once, when it is written
Text from outside your team is screened as it is stored, so every agent that reads it later gets the result without running a check of its own.
A verdict on every record and file
Clean, suspicious or blocked. Your policy decides what each verdict does, and findings can be raised or relaxed within set ranges.
Held for review when it matters
In a screened collection an untrusted write waits until its check has finished or a person has released it. Reviewers release or reject from the dashboard.
A verdict and the fields to distrust
The verdict travels with the record, over MCP and in REST responses, so the agent can tell your own content from text written by a customer or another agent.
{
"version": 1,
"source": "record_write",
"verdict": "suspicious",
"reasons": ["prompt_attack"],
"screened": true,
"provenance": "recorded",
"tainted_paths": ["/body"],
"writers": { "/body": "flux_key" }
}tainted_paths lists the fields that came from outside your team; writers says who wrote each one.
Your agent can drop a suspicious result, quote it as data, or ask a person. How trust labels work →
An injected instruction has little to reach
An agent works with a tenant token for one customer, and writes only where you switch writes on. Here a shop assistant is asked for a discount and tries to change a price; writes are off there, so nothing changes.
A shop assistant, asked by a customer for a discount, tries to change a price. Writes are off for prices, so nothing is written; it gets collection_not_writable and the list of collections it may write to: tickets and notes.
Each customer's data sits in its own space. Multi-tenancy →
Detect prompt injection before your agent reads it
Turn on checks for a collection, mark who may write, and let your agents read everything with a verdict.