Skip to content

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

fix(providers): omit complete text values containing attachment URLs

Emre Sokullu7849097a3003
diff --git a/src/chrome/src/trace/webbrain-share-outbox.js b/src/chrome/src/trace/webbrain-share-outbox.js
index 314b6e109..fae2a3b20 100644
--- a/src/chrome/src/trace/webbrain-share-outbox.js
+++ b/src/chrome/src/trace/webbrain-share-outbox.js
@@ -11,6 +11,7 @@ const MAX_OUTBOX_ITEMS = 100;
const MAX_MESSAGE_CHARS = 10_000;
const MAX_REQUEST_BUDGET = 150_000;
const MAX_RESPONSE_CHARS = 40_000;
+const BINARY_DATA_URL = /data:(image|audio|video|application|font|model)\/[^\s;,]+(?:\s*;[^,]*)?\s*,/i;
let storageQueue = Promise.resolve();
let flushQueue = Promise.resolve();
let fallbackRunCounter = 0;
@@ -34,26 +35,10 @@ function clampText(value, limit = MAX_MESSAGE_CHARS) {
function scrubText(value, limit = MAX_MESSAGE_CHARS) {
if (typeof value !== 'string') return value;
let text = value;
- // Non-canonical forms count too: uppercase scheme (DATA:...), extra
- // media-type parameters (...;charset=utf-8;base64,...). Case-insensitive.
- if (/data:/i.test(text)) {
- const headers = text.match(/data:(image|audio|video|application|font|model)\/[^,\s"'<>]+,/gi) || [];
- // Non-base64 data URLs may contain raw markup, quotes, and whitespace.
- // Their end cannot be distinguished reliably from surrounding text, so
- // omit the whole value instead of risking a partial attachment upload.
- if (headers.some(header => !/;base64,$/i.test(header))) return '[binary content omitted]';
- // Detect data URIs regardless of surrounding string or payload length:
- // even a 1x1 canvas/QR thumbnail is image bytes the UI promises to strip.
- // Newlines (but not spaces/words) are allowed inside the payload so
- // wrapped base64 is still removed while surrounding prose survives.
- text = text.replace(
- /data:(image|audio|video|application|font|model)\/[a-zA-Z0-9+.-]+(?:;[a-zA-Z0-9!#$&^_.+-]+(?:=[a-zA-Z0-9!#$&^_.+-]+)?)*;base64,[A-Za-z0-9+/=\r\n]+(?=$|[\s"'<>])/gi,
- '[embedded base64 data omitted]',
- );
- // Unhandled forms (including escaped base64) must not leave a tail of
- // bytes behind after a partial match.
- if (containsBinaryBlock(text)) return '[binary content omitted]';
- }
+ // Data URL payloads can contain markup, quotes, whitespace, and escapes,
+ // even with base64 encoding. Omit the entire value: guessing its end can
+ // leave image/file bytes behind. The media type is enough to identify it.
+ if (BINARY_DATA_URL.test(text)) return '[binary content omitted]';
// Serialized tool results can contain binaries of any size. Scrub named
// base64 fields even when their payload is below the bare-blob threshold.
if (/"base64"/i.test(text)) {
@@ -74,9 +59,7 @@ 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') {
- // Match binary media types regardless of data-URL encoding or parameters.
- return /data:/i.test(value)
- && /data:(image|audio|video|application|font|model)\/[^,\s"'<>]+,/i.test(value);
+ return BINARY_DATA_URL.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 314b6e109..fae2a3b20 100644
--- a/src/firefox/src/trace/webbrain-share-outbox.js
+++ b/src/firefox/src/trace/webbrain-share-outbox.js
@@ -11,6 +11,7 @@ const MAX_OUTBOX_ITEMS = 100;
const MAX_MESSAGE_CHARS = 10_000;
const MAX_REQUEST_BUDGET = 150_000;
const MAX_RESPONSE_CHARS = 40_000;
+const BINARY_DATA_URL = /data:(image|audio|video|application|font|model)\/[^\s;,]+(?:\s*;[^,]*)?\s*,/i;
let storageQueue = Promise.resolve();
let flushQueue = Promise.resolve();
let fallbackRunCounter = 0;
@@ -34,26 +35,10 @@ function clampText(value, limit = MAX_MESSAGE_CHARS) {
function scrubText(value, limit = MAX_MESSAGE_CHARS) {
if (typeof value !== 'string') return value;
let text = value;
- // Non-canonical forms count too: uppercase scheme (DATA:...), extra
- // media-type parameters (...;charset=utf-8;base64,...). Case-insensitive.
- if (/data:/i.test(text)) {
- const headers = text.match(/data:(image|audio|video|application|font|model)\/[^,\s"'<>]+,/gi) || [];
- // Non-base64 data URLs may contain raw markup, quotes, and whitespace.
- // Their end cannot be distinguished reliably from surrounding text, so
- // omit the whole value instead of risking a partial attachment upload.
- if (headers.some(header => !/;base64,$/i.test(header))) return '[binary content omitted]';
- // Detect data URIs regardless of surrounding string or payload length:
- // even a 1x1 canvas/QR thumbnail is image bytes the UI promises to strip.
- // Newlines (but not spaces/words) are allowed inside the payload so
- // wrapped base64 is still removed while surrounding prose survives.
- text = text.replace(
- /data:(image|audio|video|application|font|model)\/[a-zA-Z0-9+.-]+(?:;[a-zA-Z0-9!#$&^_.+-]+(?:=[a-zA-Z0-9!#$&^_.+-]+)?)*;base64,[A-Za-z0-9+/=\r\n]+(?=$|[\s"'<>])/gi,
- '[embedded base64 data omitted]',
- );
- // Unhandled forms (including escaped base64) must not leave a tail of
- // bytes behind after a partial match.
- if (containsBinaryBlock(text)) return '[binary content omitted]';
- }
+ // Data URL payloads can contain markup, quotes, whitespace, and escapes,
+ // even with base64 encoding. Omit the entire value: guessing its end can
+ // leave image/file bytes behind. The media type is enough to identify it.
+ if (BINARY_DATA_URL.test(text)) return '[binary content omitted]';
// Serialized tool results can contain binaries of any size. Scrub named
// base64 fields even when their payload is below the bare-blob threshold.
if (/"base64"/i.test(text)) {
@@ -74,9 +59,7 @@ 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') {
- // Match binary media types regardless of data-URL encoding or parameters.
- return /data:/i.test(value)
- && /data:(image|audio|video|application|font|model)\/[^,\s"'<>]+,/i.test(value);
+ return BINARY_DATA_URL.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 442fc38bf..4905f09e0 100644
--- a/test/run.js
+++ b/test/run.js
@@ -12578,7 +12578,7 @@ test('Share-for-research scrub removes embedded data URIs and keeps repeated-ans
const serialized = JSON.stringify(item);
assert.equal(serialized.includes('iVBORw0KGgoAAAANSUhEUg'), false, `${label}: tool-result data URI escaped the scrub`);
assert.equal(serialized.includes('ABCD1234abcd'), false, `${label}: array string data URI escaped the scrub`);
- assert.match(serialized, /embedded base64 data omitted/, `${label}: data-URI placeholder missing`);
+ assert.match(serialized, /binary content omitted/, `${label}: data-URI placeholder missing`);
// Tiny thumbnails in short strings must be scrubbed too, not just long payloads.
const tiny = outbox.buildShareGenerationItem({
runId: `run-share-tiny-${label}`,
@@ -12718,7 +12718,7 @@ test('Share-for-research scrubs wrapped bare base64 without counting line breaks
}
});
-test('Share-for-research strips non-base64 attachment data URLs in text and nested content', () => {
+test('Share-for-research omits text and nested content containing attachment data URLs', () => {
for (const [label, outbox] of [['chrome', SHARE_OUTBOX_CH], ['firefox', SHARE_OUTBOX_FX]]) {
for (const dataUrl of [
'data:image/svg+xml,%3Csvg%20xmlns=%22http://www.w3.org/2000/svg%22%3E%3C/svg%3E',
@@ -12733,6 +12733,10 @@ test('Share-for-research strips non-base64 attachment data URLs in text and nest
'data:image/svg+xml,<?xml version="1.0"?>\n<svg><rect width="10" height="10"/></svg>',
'data:image/svg+xml,<svg viewBox="0 0 10 10"/>',
'data:image/svg+xml,<svg><text>INCOMPLETE_IMAGE',
+ 'data:image/png;base64,QU JD',
+ 'data:image/png;base64,QU\tJD',
+ 'data:image/png;name="sample";base64,QUJD',
+ 'data:image/png;name="sample,file";base64,QUJD',
]) {
const content = `before "${dataUrl}" after`;
const expected = '[binary content omitted]';