Council of High Intelligence distributes prompt files (skills + agent personas) and shell scripts that users install into their AI coding tools (Claude Code, Codex, Gemini CLI). The threat model is therefore supply-chain shaped:

• Prompt injection — a malicious change to SKILL*.md or agents/*.md could smuggle instructions into a user's AI assistant.
• Shell execution — install.sh and scripts/*.sh run on user machines; the coordinator protocol dispatches provider CLIs via shell.
• Credential handling — provider API keys (e.g. NVIDIA_API_KEY, CURSOR_API_KEY) are resolved from environment variables at runtime and must never be inlined, logged, or placed in process argv.

Please do not open a public issue for security-sensitive reports. Instead use GitHub private vulnerability reporting.

You can expect an initial response within 7 days. Verified fixes are credited in the changelog unless you prefer otherwise.

• No curl | bash install path — users clone the repo and run install.sh locally.
• No network calls at install time; provider detection only probes locally installed CLIs and, when an API key is present, the provider's /models endpoint.
• API keys are passed to curl via process substitution, never as argv.
• All contributed prompt/script changes are reviewed for embedded instructions, obfuscation (base64, hex blobs), and unexpected URLs before merge.

Only the latest release is supported with security fixes.