File Checks
Every uploaded file is checked before anyone can read it. A refused file is never delivered and its bytes are purged; the uploader sees rejected with a public reason that does not name the rule.
At a glance
| Check | agent_ready | storage_only |
|---|---|---|
| Malware scan before FoxNose reads the file | Yes | Yes |
| Format identified from content | Yes | Common formats: images, PDFs, Office files, archives, web pages, audio, video, programs |
| Content contradicting the declared type or extension refused | Yes | Yes, for the formats above |
| Encrypted and password-protected files refused | Yes | Archives, Office and OpenDocument files |
| Damaged files and files joined to another file refused | Yes | ZIP-based files, RAR and TAR |
| Macros, embedded objects and linked external content refused | Word, OpenDocument and RTF | No, stored as is |
| Rendering in an isolated sandbox | Yes | No |
| Visible text only, hidden text reported | Yes | No text is read |
| Prompt-attack, sexual-text and PII screens | Yes | No |
| Image moderation | Yes | No |
| Active content in PDFs counted | Yes | No |
| Safe copy for people | Yes | No |
| Agents receive | Visible text, verdict, flags | Metadata only |
In both modes a file of a kind the policy or upload link does not allow is refused as unsupported_format, and a file over the size limit as too_large. Each upload needs a single-use grant that expires within minutes. If a later scan finds a published file infected, it is withdrawn from search and agents and its bytes are purged.
Agent-ready collections
- Format. Identified from the bytes and validated before anything renders it. A damaged file or one with another file joined to it is refused, never repaired. Encrypted files, animated or multi-image files, and documents with macros, embedded objects, remote templates or fields that fetch other files are refused. Hyperlinks are fine. See Supported inputs.
- Rendering. Pages are rendered in a disposable sandbox with no network access. All checks below read the rendering, not the file.
- Text. OCR reads only what is drawn. Hidden text (layers, white-on-white, tiny fonts, off-page text, comments, tracked changes) is not delivered and is reported as
hidden_text. Text too faint or small to read is redacted (low_legibility_text). Invisible and direction-changing characters are stripped (unicode_smuggling). - Screens. Visible text is screened for prompt attacks, sexual content and personal data. Screens report; they never change the text.
- Moderation. Every page is checked for unsafe imagery. Explicit imagery is blocked by default.
- Active content. Scripts, launch actions and embedded files in PDFs are counted and reported, never run.
The results are findings, and the policy turns them into a verdict. People get a safe copy of every file, and original_access can withhold the original until the file is clean.
Storage-only collections
A storage_only collection scans and stores files of any type without processing their content. Agents receive metadata only. A file is stored when:
- the malware scan is clean;
- its kind is allowed:
archivefor archives,imagewhen the declared MIME type is an image type,documentotherwise; - its content does not contradict its declared type or extension, such as a program or web page named
invoice.pdf; - an archive passes the archive rules;
- an Office or OpenDocument file is not password-protected;
- a ZIP-based file is well-formed and not joined to another file.
Spreadsheets, presentations and .doc files are stored as they are, unread. People can download the original once the scan is clean, unless original_access is disabled. There is no safe copy, so when_clean is not available.
Treat every original downloaded from a storage_only collection as a file from an unknown sender: open it only with software you trust, and do not run it.
Archives
Archives are accepted only in storage_only collections, recognised by content, and stored with the kind archive. Allowing document also allows archive. FoxNose scans archives for malware but never extracts or indexes them.
| Accepted | Refused |
|---|---|
| ZIP | Encrypted archives |
| RAR 4 and 5, single volume | 7z |
| TAR, uncompressed | Compressed streams: .gz, .tgz, .bz2, .xz and their .tar.* forms |
| Multi-volume and split archives | |
| Archives containing another archive | |
| More than 10,000 entries (for RAR, internal blocks) | |
| Damaged, truncated, or joined to another file |
A refused archive answers unsupported_format. Upload a compressed TAR uncompressed or as a ZIP.
What the checks do not do
- Detectors give signals, not guarantees. Use the verdict and findings as inputs to your agent's own safeguards. See Trust model.
- Visible prompt injection is delivered, inside the untrusted envelope. The prompt-attack screen flags it but never removes it.
- A clean malware scan finds known threats; it does not prove a file is safe to open.
- Uploaders are not told why a file was refused. Your team sees the verdict and findings in the dashboard and the Management API.
Related
- Agent-Ready Files — what agents receive, verdicts, findings and policy.
- Files Limits — sizes, page counts, rates and retention.
- File content over MCP — the fields an agent receives.