Security context

What an agent needs to avoid regressing past fixes and find the next vuln in this repo.

browseros-ai/browseros

              main @ a8e5b78
            

Fixes

CVEs

CRITICAL

Peak severity

36.0%

Coverage

Highlights

Path Traversal: 11 prior fixes. Scrutinize any change in this area.

packages/browseros-agent/apps/server/src/api/routes/mcp.ts: most-fixed (2 issues). Treat as high-risk during review.

22 high-severity fixes in this history; regressions here are high-impact.

Recurring patterns

The bug types that recur here, drawn from past fixes, not open vulnerabilities.

Path Traversal: The Model Context Protocol (MCP) and agent-driven filesystem tools repeatedly fail to confine file operations to safe workspaces, allowing remote agents to read and write arbitrary files outside the execution sandbox. This has been addressed across multiple components using canonicalization and prefix verification.

Auth Bypass: ACL pattern matching and agent tool permission gates have been historically weak, permitting agents to bypass execution restrictions using multi-segment paths or semantic nuances. Robust NLP scoring and signature validation represent primary mitigations.

Command Injection: GitHub Actions workflows interpolate untrusted input configurations and version strings directly into inline shell scripts instead of utilizing safe environment variables, leading to potential command injection in the build environment.