agent-browser

SUSPICIOUS. The skill’s capabilities broadly match browser automation, but its trust footprint is larger than a simple local Playwright helper: it requires transitive skill installation, relies on an external CLI/service, and forwards sensitive browsing, login, proxy, upload, screenshot, cookie, and video data through that channel. Official provenance signals reduce concern about outright malware, but the combination of external dependency trust, credential forwarding, arbitrary-site browsing, and data-capture features makes this a medium-high security risk skill.

The provided fragment contains a clear high-sensitivity capability: it automates login and then harvests authenticated browser cookies via injected JavaScript (document.cookie), parses the results, and echoes the cookies to stdout. Even without visible malware behaviors beyond browser automation, this is strongly aligned with session/credential harvesting patterns and should be treated as a significant supply-chain/security risk unless strictly confined to an isolated, controlled test environment with strict access controls and output redaction. Additional context (full login implementation, whether cookies are actually emitted, and how outputs are handled in the consuming project) would be needed to reduce uncertainty about malicious intent.