The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted telemetry and configuration data from target web pages via browser globals such as window.__GAME_DIAGNOSTICS__ and window.__GAME_PLAYTEST__. This data is used to inform test logic and is passed to an external API for AI-based decision making. The ingestion of these arbitrary inputs from external web pages presents a surface for indirect prompt injection, where a malicious page could attempt to influence the agent's actions or decisions.
Ingestion points: Browser-based diagnostics are read using vg playtest eval within the scripts/scripted-playtest.mjs and scripts/lib/harness.mjs scripts.
Boundary markers: The skill does not implement explicit boundary markers or sanitization for the data retrieved from the browser context.
Capability inventory: The skill can execute shell commands via vg and node, and performs network communication with the vibedgames API.
Sanitization: There is no evidence of filtering or escaping logic for the content retrieved from the target page's globals before it is processed or transmitted.
[DYNAMIC_EXECUTION]: The skill makes extensive use of JavaScript evaluation (eval) to interact with the internal state of the games being tested. This includes reading telemetry, triggering test hooks, and dispatching synthetic input events to the browser. This is a core functional requirement for the skill's automation capabilities.
[COMMAND_EXECUTION]: The skill executes external processes, including the vg CLI tool and internal Node.js scripts, to manage browser sessions and orchestrate testing workflows.
[EXTERNAL_DOWNLOADS]: The skill's documentation specifies that the vg utility automatically installs the agent-browser tool and its required browser binaries from Vercel Labs' official repositories during its initial setup. Vercel Labs is a well-known organization.
[DATA_EXFILTRATION]: The autonomous playtest feature (vg playtest run) transmits live game diagnostics from the browser to the vibedgames API to receive model-generated movement decisions. This data transfer is documented as a functional requirement for the AI-assisted testing loop.

playtest