fix(agent): harden _extract_start_url URL gating — skip local paths (incl. quoted) and match file extensions on the path, not as a substring (#4794)#4983
Open
r266-tech wants to merge 3 commits into
Conversation
…_start_url (browser-use#4794) A local file path in the task (e.g. an uploaded /app/x_capabilities.html) matched the domain regex and was force-navigated as the agent's first action. Skip URL candidates whose whitespace-delimited token is a local filesystem path (/, ~/, ./, ../, or a drive). Fixes both the agent and rust _extract_start_url copies; adds a regression test. Avoids the extension-only approach, which would regress legit .html/.htm URLs.
Contributor
There was a problem hiding this comment.
1 issue found across 3 files
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="browser_use/agent/service.py">
<violation number="1" location="browser_use/agent/service.py:2391">
P2: Quoted/parenthesized local file paths bypass the local-path filter because surrounding delimiters are included in `local_path_token`, causing the anchored regex to fail.</violation>
</file>
Reply with feedback, questions, or to request a fix.
Fix all with cubic | Re-trigger cubic
…ract_start_url
The excluded-extension filter tested f'.{ext}' as a SUBSTRING of the whole
URL, so everyday sites whose host/path merely contains a short extension
token were wrongly dropped from auto-navigation (e.g. '.py' in
docs.python.org, '.doc' in my.docs.google.com, '.js'/'.css' hosts). Decide
exclusion from the final path segment instead (scheme/query/fragment
stripped, trailing slash removed, percent-decoded, path-params split), so
genuine downloadable-file URLs (report.pdf, data.json;v=1, archive.tar.gz)
still drop. Also strip leading quote/paren delimiters before the local-path
guard so quoted local paths are skipped too. Mirrored in agent + rust copies;
regression tests exercise both implementations.
Contributor
There was a problem hiding this comment.
1 issue found across 3 files (changes from recent commits).
Reply with feedback, questions, or to request a fix.
Fix all with cubic | Re-trigger cubic
…_extract_start_url
The local-path guard lstrip()'d quote/paren delimiters ('"(<[{) but not the
Markdown backtick, so a backtick-wrapped local path (e.g. `/app/x.html`) kept
its leading backtick, the anchored filesystem-path regex failed, and since
'html' is intentionally not an excluded extension the path was auto-navigated
as a bare URL. Add ` to the stripped delimiter set in both agent + rust copies;
regression tests cover backtick-wrapped local paths in both implementations.
Contributor
Author
|
Good catch — addressed in 4267497. The local-path guard's delimiter strip set covered |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #4794.
_extract_start_url(used whendirectly_open_url=True, the default) scans the task text and force-injects an initialnavigateaction before the LLM acts. This PR hardens that URL-gating against two false-decision classes, in both the agent and rust copies, with regression tests covering both implementations.1. Local filesystem paths were force-navigated (#4794).
A local file path in the task — e.g. an uploaded
/app/x_capabilities.html— matched the domain regex (the domain char class[a-zA-Z0-9-]excludes_, sox_capabilities.htmlis captured ascapabilities.html) and was navigated to as the agent's first action, then blocked by the SecurityWatchdog, derailing the run. Such candidates are now skipped. Quoted/parenthesised local paths (e.g."/app/x.html") are skipped too — leading quote/paren delimiters are stripped before the local-path check.2. Everyday sites were wrongly excluded as "files".
The excluded-extension filter tested
f'.{ext}'as a substring of the whole URL, so any host/path merely containing a short extension token was dropped from auto-navigation:docs.python.org(.py),www.python.org(.py),my.docs.google.com(.doc), and any.css/.js/.mdhost. Exclusion is now decided from the final path segment (scheme/query/fragment stripped, trailing slash removed, percent-decoded,;path-params split), so genuine downloadable-file URLs still drop (report.pdf,report.pdf/,data.json;v=1,report%2Epdf,archive.tar.gz) while real pages are kept (index.html,example.com/report.pdf/view— a page, not a file).Deliberate scope note: exclusion keys on the path, not the query string. A URL like
example.com/download?file=report.pdf(a download endpoint) orexample.com/view?doc=report.pdf(a viewer page) is kept — the old substring check dropped both, which over-excluded navigable viewer/search pages. Navigation policy is still enforced downstream by the SecurityWatchdog.