PromptShieldCONTENT INSPECTOR
Connecting

YOUR FIRST LINE OF REVIEW

Inspect untrusted content.

Find the instructions hiding inside your email and documents.
Understand the risk before passing anything to your AI.

Protected modeInspect first. Generate after.

01 Source

Checking your session…

02 Inspection

READY

AWAITING YOUR SOURCE

Your findings belong here.

Add a file or paste an email to see its risk score,
the evidence behind it, and what was inspected.

Risk score Scan coverage Source preview

A clear result starts with a complete inspection.

Source content is data. It cannot authorize a new task.

BEFORE YOU RELY ON A RESULT

Scope & limitations

PromptShield inspects untrusted content before an optional local model task. In protected mode, flagged sources and incomplete scans are held. A clean result means no indicators were found in the supported content.

Email & HTML
Decoded email headers and bodies; static visible text, comments, hidden text, attributes, and active-source evidence. Attachments and remote resources are omitted. Scripts never run.
PDF
Extractable text and supported metadata, annotations, and bookmarks. No OCR. Images, encrypted documents, forms, and embedded files can leave coverage incomplete.
Limits
Email and HTML: 1 MiB. PDF: 5 MiB, 20 pages. Up to 200,000 extracted characters. Decoding is bounded; exceeding a budget holds the source.
Protection
Deterministic boundary checks, bounded decoding, and the preserved local classifier with semantic guards. Output is validated before it is shown. No real actions are executed.
Evidence
This new protected policy has engineering regression tests. Existing research measurements describe earlier frozen policies. They do not establish this version's accuracy on arbitrary files. Detection can miss attacks and flag ordinary text.
Privacy
Inputs and results are processed on this host. The application does not save their content in logs. Source content and results stay in this page until you clear it or close it. Your theme preference is remembered on this device. An exported report contains the displayed source preview and findings; export it only when you intend to save that content. Local runtime memory is not guaranteed to be securely erased.