The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill integrates with OpenAI's Decisions API to classify user intents. The implementation in src/core/providers/decisions.ts uses process.env.OPENAI_API_KEY and sends requests to https://api.openai.com/v1/decisions. The documentation in SKILL.md and docs/wiring.md explicitly warns against exposing this key to the client (e.g., via NEXT_PUBLIC_* variables) and recommends standard secret management practices.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided search queries at runtime. It implements mitigation strategies in src/core/normalize.ts, including length limits and rejection of control characters, backslashes, and path traversal sequences (..). Furthermore, src/core/resolve.ts uses JSON quoting to ensure user input cannot pose as instructions to the underlying LLM.
[COMMAND_EXECUTION]: During the setup phase, the skill instructs the agent to copy its own source files into the project directory using cp -R. These files are part of the skill package itself and do not involve downloading or executing untrusted remote code.
[PROMPT_INJECTION]: While a static detector flagged instructional language in SKILL.md regarding automatic configuration choices, an adversarial review confirms these are benign developer instructions for optimizing the tool's performance and cost during installation, not attempts to bypass safety filters or conceal malicious behavior.

intentbar-setup