The static command findings are mostly false positives from markdown command examples, not Ruby backtick execution. The skill is still medium risk because it intentionally allows browser automation, arbitrary Playwright page code, network navigation, storage-state files, persistent profiles, and trace artifacts that can expose sensitive session data. No prompt injection attempt or confirmed malicious exfiltration path was found.
The skill documents saving, loading, listing, and editing cookies, localStorage, sessionStorage, and authentication state files. This is legitimate for testing, but saved state can contain session tokens and must be protected.
The lines explicitly show state-save, state-load, cookie commands, and security notes about auth tokens. The context is legitimate automation, so the risk is exposure through misuse rather than confirmed malicious intent.
Arbitrary Playwright Code Execution in Browser Context
The skill supports run-code and eval commands that execute user supplied Playwright or JavaScript snippets in the active browser session. This enables advanced testing, but it can read page content, grant browser permissions, and manipulate storage.
The commands are directly documented and the examples show page evaluation and permission changes. The capability is expected for a Playwright CLI skill, but it increases impact if misused.
Trace and Video Artifacts May Capture Sensitive Data
Tracing records DOM snapshots, screenshots, network requests, headers, bodies, and console logs. Video recording and traces can preserve secrets typed during automated flows.
The reference states that traces capture network bodies, DOM snapshots, screenshots, and checkout form inputs. This is strong evidence of sensitive artifact risk without evidence of automatic exfiltration.
The skill supports persistent sessions and custom profile directories. Persistent profiles can keep cookies, local storage, and other browser data beyond the immediate automation task.
The documented commands explicitly create persistent sessions and provide cleanup commands. The risk is manageable, but users must understand retention and delete data after sensitive workflows.
Static External Command Matches Are Mostly Documentation Examples
The many Ruby or shell backtick detections occur in fenced markdown blocks that show playwright-cli usage. They are not Ruby backtick execution in source code, although the skill intentionally allows the playwright-cli command.
The cited locations are markdown command examples, not executable Ruby code. The allowed Bash wrapper is real, but the specific static Ruby backtick label is a false positive.
The hardcoded URL findings use documentation domains such as example.com and playwright.dev. I did not find evidence that these URLs receive secrets or hidden telemetry.
The URL examples are clearly part of user-facing command documentation and test workflows. No evidence found of a fixed third-party collection endpoint.
No reviewed file contained suspicious override language, fake system instructions, pre-approval claims, or requests to skip analysis.
Targeted review of the skill and references found no prompt injection phrases. The confidence is high, but this finding is limited to the files included in the audit.