← All engineering notes

Enterprise AI

Account for attachments your AI controls cannot inspect

Inventory attachment types and choose a deliberate response when a control has no coverage.

An attachment is not automatically visible to every control in an AI request path. A PDF may have an extractable text layer; an image, audio file, encrypted archive or remote URL may not. Record coverage per attachment type and decide whether to reject, isolate, scan elsewhere or accept the gap.

The NIST AI Risk Management Framework frames AI risk management as a lifecycle activity. Use that approach for input coverage: inventory the actual file types and channels your application accepts, then test the controls against each one. A single “files supported” statement does not describe what any detector read.

Make coverage visible in the request decision

Input Example limitation Suitable response depends on risk
Text PDF Extraction can omit rendered or hidden content Inspect extracted text and flag incomplete extraction
Image or scan Text detector cannot see pixels without OCR or vision processing Reject for sensitive workflows, or use a separate reviewed path
Audio or video Transcript and media can disagree Restrict supported formats or state the uninspected coverage
Link to remote file Destination can change after admission Fetch through a controlled service and record the resolved source

OWASP’s prompt-injection guidance identifies hidden text and multimodal inputs as injection paths. It recommends treating remote content as untrusted and checking whether extracted text matches what a human can see. This reduces risk; it does not make every file safe.

Choose a policy for no coverage

Use one of three explicit outcomes when a required control cannot inspect an attachment:

  1. Refuse: return a clear reason and direct the user to an approved format.
  2. Quarantine: store or process the item in a specialist workflow before the model sees it.
  3. Allow with recorded gap: reserve this for low-risk workflows where the owner accepts the limitation.

Do not silently pass the file through while recording an “inspected” verdict. That turns missing coverage into false evidence. An allow-with-gap path should say which detector did not run, why, and which policy approved it.

For a hypothetical HR assistant, a scanned résumé may be eligible for a malware scan and manual review but ineligible for automated retrieval into a model until OCR quality and hidden-content tests pass. A financial-support assistant may instead reject the scan and ask for a structured form. The correct choice depends on the data and the action the model can trigger.

Alectura reads the text layer of PDF input for its guardrails. It records image, audio and video input as uninspected, and a policy can refuse such content with attachment_uninspectable. Its documentation describes that limitation. Review the coverage table with your security owner, then test one internal application using representative but non-sensitive files.

Repeat the test when the application adds a new provider, parser or upload surface. Coverage can change even when the user-facing file picker looks the same. Keep the result with the release record so support staff can explain whether a refusal protects the user or whether an accepted attachment passed through a known control gap.

Give product teams a supported alternative for rejected files. A clear error can offer an approved text format, an internal scanning route, or a human review queue. That reduces pressure to add a broad exception during a customer escalation. The alternative still needs its own data-flow review before it becomes the easy path around an inspection limit.

Test coverage with each release

Do not make the coverage matrix a one-time security questionnaire. Put it beside the application’s acceptance tests and release checklist. A test can assert that a scan produces an uninspected decision, that a strict policy returns the documented refusal, and that logs name the attachment class without storing the file. When a parser upgrade begins to read another format, review the changed coverage before you advertise it as protection.

Sources checked 14 November 2026. Provider capabilities and terms can change; check the linked documentation for your configuration.

Keep reading

Trace an internal AI incident without collecting every employee prompt ↗Build a secrets-redaction pipeline before AI requests leave your boundary ↗Validate LLM tool arguments before your application runs them ↗