Fixed media transcripts reaching the AI without the prompt injection filter (#1083)

app/MindWork AI Studio/Tools/Media/PersistentMediaTranscript.cs

@@ -3,6 +3,8 @@ using System.Text;
 using AIStudio.Settings;
 using AIStudio.Tools.PluginSystem;
 using AIStudio.Tools.Rust;
+using AIStudio.Tools.Security;
+using AIStudio.Tools.Services;
 
 namespace AIStudio.Tools.Media;
 
@@ -155,6 +157,29 @@ internal static class PersistentMediaTranscript
         }
     }
 
+    /// <summary>
+    /// Reads a transcript for an AI, through the runtime like every document.
+    /// </summary>
+    /// <remarks>
+    /// The runtime filters prompt injections while it reads, so a transcript must never reach an AI
+    /// by any other way, such as reading the file directly. What the filter removed is reported under
+    /// the recording: the transcript is a file AI Studio made, while the recording is the file the
+    /// user chose.
+    /// </remarks>
+    /// <param name="rustService">The runtime service.</param>
+    /// <param name="transcript">A transcript with the status READY.</param>
+    /// <param name="mediaPath">The audio or video file the transcript belongs to.</param>
+    /// <param name="reportingScope">The action the transcript is read for, or null when it is read for none.</param>
+    /// <param name="token">The cancellation token.</param>
+    /// <returns>The filtered text of the transcript, or why it could not be read.</returns>
+    public static Task<FileExtractionResult> ReadTranscriptAsync(RustService rustService, PersistentMediaTranscriptResult transcript, string mediaPath, PromptInjectionReportingScope? reportingScope, CancellationToken token)
+    {
+        if (transcript.Status is not PersistentMediaTranscriptStatus.READY)
+            throw new ArgumentException($"Only a ready transcript can be read, but its status is {transcript.Status}.", nameof(transcript));
+
+        return rustService.ReadArbitraryFileData(transcript.TranscriptPath, int.MaxValue, reportAs: PromptInjectionSource.FileContent(mediaPath), reportingScope: reportingScope, token: token);
+    }
+
     private static PersistentMediaTranscriptResult SourceChanged(MediaTranscriptSlot slot) => PersistentMediaTranscriptResult.Unavailable(PersistentMediaTranscriptStatus.SOURCE_CHANGED, slot, TB("The media file changed while it was being transcribed, so its transcript was not kept."));
 
     /// <summary>

app/MindWork AI Studio/Tools/Media/PersistentMediaTranscriptResult.cs

@@ -8,7 +8,9 @@ namespace AIStudio.Tools.Media;
 /// <remarks>
 /// The transcript itself is not part of it: whoever needs the text reads the file at
 /// TranscriptPath, so a transcript reaches the AI by the same path no matter whether it was just
-/// made or stored long ago.
+/// made or stored long ago. That read always goes through the runtime, which filters prompt
+/// injections: PersistentMediaTranscript.ReadTranscriptAsync reads a transcript as a whole, and a
+/// data source reads it in pieces for its index.
 /// </remarks>
 /// <param name="Status">How the request ended.</param>
 /// <param name="TranscriptPath">The path of the transcript; empty when the place could not be determined.</param>