The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/parse-claude-logs.py accesses and parses log files located in ~/.claude/logs/. These logs contain full conversation histories, request/response metadata, and potentially sensitive debug information which are exposed to the tool for metric extraction.
[CREDENTIALS_UNSAFE]: scripts/benchmark.py and the orchestrator instructions in references/orchestrator-patterns.md programmatically retrieve sensitive authentication credentials, such as ANTHROPIC_API_KEY, OPENAI_API_KEY, and AWS_ACCESS_KEY_ID, from environment variables to authorize benchmark requests.
[COMMAND_EXECUTION]: The orchestrator workflow patterns described in references/orchestrator-patterns.md instruct the subagent to execute shell commands, including date for high-precision timestamps and echo $ANTHROPIC_BASE_URL to verify configuration, which are used to coordinate the benchmarking sequence.
[EXTERNAL_DOWNLOADS]: The benchmark tool initiates network connections to well-known AI service providers, including Anthropic (api.anthropic.com), OpenAI (api.openai.com), and Google Gemini. These are recognized as trusted services, and the connections are consistent with the skill's stated purpose.
[INDIRECT_PROMPT_INJECTION]: The skill processes task-specific prompts (e.g., in references/task-prompts.md) and user-provided inputs without explicit boundary markers or sanitization. When processed by subagents with capabilities to write files and access networks, this creates a vulnerability surface for indirect prompt injection, although the risk is low due to the tool's specialized use case.

llm-api-benchmark