The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the user to run a sudo security add-trusted-cert command to modify the system's root trust store. This is a high-risk operation that compromises the machine's security by allowing the proxy to decrypt HTTPS traffic. Additionally, the latency_inject.lua script uses os.execute to invoke the system sleep command, which provides a pathway for arbitrary command execution.\n- [CREDENTIALS_UNSAFE]: The log_to_jsonl.lua and capture_for_eval.lua scripts record full HTTP response headers and bodies to local files (~/.proxelar/traffic.jsonl). These logs often contain sensitive secrets such as session cookies, Authorization headers (API keys), and personal data in plain text.\n- [EXTERNAL_DOWNLOADS]: The skill instructions direct users to install third-party software from the internet via brew install proxelar. This introduces unverified external binaries into the execution environment.\n- [DATA_EXFILTRATION]: The skill is specifically designed to target and record traffic to major LLM providers (Anthropic, OpenAI, Gemini), harvesting prompt and response data that may contain proprietary or confidential information.\n- [PROMPT_INJECTION]: The skill creates a significant surface for indirect prompt injection (Category 8).\n
Ingestion points: Intercepts all incoming and outgoing network traffic via on_request and on_response hooks in multiple scripts (log_to_jsonl.lua, capture_for_eval.lua, etc.).\n
Boundary markers: None identified; traffic is processed as raw strings.\n
Capability inventory: Ability to execute shell commands via os.execute, write to the filesystem via io.open, and modify network traffic content before it reaches the agent.\n
Sanitization: None; the scripts perform string matching and manipulation on raw, untrusted network data.

proxelar