Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

fix(providers): strip non-base64 attachment data URLs

Emre Sokullu00ee165bffe4
diff --git a/src/chrome/src/trace/webbrain-share-outbox.js b/src/chrome/src/trace/webbrain-share-outbox.js
index 975ba697d..e3713fb08 100644
--- a/src/chrome/src/trace/webbrain-share-outbox.js
+++ b/src/chrome/src/trace/webbrain-share-outbox.js
@@ -45,6 +45,13 @@ function scrubText(value, limit = MAX_MESSAGE_CHARS) {
/data:(image|audio|video|application|font|model)\/[a-zA-Z0-9+.-]+(?:;[a-zA-Z0-9!#$&^_.+-]+(?:=[a-zA-Z0-9!#$&^_.+-]+)?)*;base64,[A-Za-z0-9+/=\r\n]+/gi,
'[embedded base64 data omitted]',
);
+ // Data URLs can also carry percent-encoded or unencoded bytes (SVG,
+ // PDFs, etc.). Their media type identifies the attachment regardless of
+ // encoding. Stop at text/JSON delimiters to preserve surrounding prose.
+ text = text.replace(
+ /data:(image|audio|video|application|font|model)\/[^,\s"'<>]+,[^\s"'<>]*/gi,
+ '[embedded data omitted]',
+ );
}
// Serialized tool results can contain binaries of any size. Scrub named
// base64 fields even when their payload is below the bare-blob threshold.
@@ -66,9 +73,9 @@ function containsBinaryBlock(value) {
// screenshot bytes echoed into a tool result). Check anywhere in the
// string, at any length: even tiny thumbnails are image bytes.
if (typeof value === 'string') {
- // Case-insensitive with optional media-type parameters, matching scrubText.
+ // Match binary media types regardless of data-URL encoding or parameters.
return /data:/i.test(value)
- && /data:(image|audio|video|application)\/[^,]+;base64,/i.test(value);
+ && /data:(image|audio|video|application|font|model)\/[^,\s"'<>]+,/i.test(value);
}
if (!value || typeof value !== 'object') return false;
const type = typeof value.type === 'string' ? value.type.toLowerCase() : '';
diff --git a/src/firefox/src/trace/webbrain-share-outbox.js b/src/firefox/src/trace/webbrain-share-outbox.js
index 975ba697d..e3713fb08 100644
--- a/src/firefox/src/trace/webbrain-share-outbox.js
+++ b/src/firefox/src/trace/webbrain-share-outbox.js
@@ -45,6 +45,13 @@ function scrubText(value, limit = MAX_MESSAGE_CHARS) {
/data:(image|audio|video|application|font|model)\/[a-zA-Z0-9+.-]+(?:;[a-zA-Z0-9!#$&^_.+-]+(?:=[a-zA-Z0-9!#$&^_.+-]+)?)*;base64,[A-Za-z0-9+/=\r\n]+/gi,
'[embedded base64 data omitted]',
);
+ // Data URLs can also carry percent-encoded or unencoded bytes (SVG,
+ // PDFs, etc.). Their media type identifies the attachment regardless of
+ // encoding. Stop at text/JSON delimiters to preserve surrounding prose.
+ text = text.replace(
+ /data:(image|audio|video|application|font|model)\/[^,\s"'<>]+,[^\s"'<>]*/gi,
+ '[embedded data omitted]',
+ );
}
// Serialized tool results can contain binaries of any size. Scrub named
// base64 fields even when their payload is below the bare-blob threshold.
@@ -66,9 +73,9 @@ function containsBinaryBlock(value) {
// screenshot bytes echoed into a tool result). Check anywhere in the
// string, at any length: even tiny thumbnails are image bytes.
if (typeof value === 'string') {
- // Case-insensitive with optional media-type parameters, matching scrubText.
+ // Match binary media types regardless of data-URL encoding or parameters.
return /data:/i.test(value)
- && /data:(image|audio|video|application)\/[^,]+;base64,/i.test(value);
+ && /data:(image|audio|video|application|font|model)\/[^,\s"'<>]+,/i.test(value);
}
if (!value || typeof value !== 'object') return false;
const type = typeof value.type === 'string' ? value.type.toLowerCase() : '';
diff --git a/test/run.js b/test/run.js
index 91ca6fd88..4f86bff17 100644
--- a/test/run.js
+++ b/test/run.js
@@ -12718,6 +12718,43 @@ test('Share-for-research scrubs wrapped bare base64 without counting line breaks
}
});
+test('Share-for-research strips non-base64 attachment data URLs in text and nested content', () => {
+ for (const [label, outbox] of [['chrome', SHARE_OUTBOX_CH], ['firefox', SHARE_OUTBOX_FX]]) {
+ for (const dataUrl of [
+ 'data:image/svg+xml,%3Csvg%20xmlns=%22http://www.w3.org/2000/svg%22%3E%3C/svg%3E',
+ 'DATA:IMAGE/SVG+XML;charset=utf-8,%3Csvg%3E%3C/svg%3E',
+ 'data:application/pdf,%25PDF-1.7%0Aendobj',
+ 'data:application/octet-stream,ABC',
+ 'data:audio/wav,%52%49%46%46',
+ 'data:video/mp4,%00%00%00%20ftyp',
+ 'data:font/woff,%77%4F%46%46',
+ 'data:model/gltf+json,%7B%22asset%22:%7B%7D%7D',
+ ]) {
+ const content = `before "${dataUrl}" after`;
+ const expected = 'before "[embedded data omitted]" after';
+ const item = outbox.buildShareGenerationItem({
+ finalContent: content,
+ messages: [
+ { role: 'tool', content },
+ { role: 'tool', content: ['caption', dataUrl, { type: 'text', text: dataUrl }] },
+ { role: 'user', content: [{ type: 'text', text: 'Read this.' }, { type: 'other', source: { url: dataUrl } }] },
+ ],
+ });
+ assert.equal(item.request[0].content, expected, `${label}: ${dataUrl}`);
+ assert.deepEqual(item.request[1].content, ['caption', '[embedded data omitted]']);
+ assert.deepEqual(item.request[2].content, [{ type: 'text', text: 'Read this.' }]);
+ assert.equal(item.response.content, expected);
+ }
+ const textUrl = 'data:text/plain,ordinary%20text';
+ const item = outbox.buildShareGenerationItem({
+ finalContent: textUrl,
+ messages: [{ role: 'user', content: textUrl }],
+ });
+ assert.equal(item.request[0].content, textUrl, `${label}: plain text URL changed`);
+ assert.equal(item.response.content, textUrl);
+ }
+});
+
test('Share-for-research caps count wrapper messages and serialized overhead', () => {
const messages = [{ role: 'system', content: 'SYS' }];
for (let i = 0; i < 250; i++) messages.push({ role: 'user', content: `cap${i}-` + 'word '.repeat(150) });