Only the latest published npm release (@proactive-agent/mcp) is supported for security fixes. Older versions should be upgraded as soon as possible.

ProactiveAgent is designed with a local-first, anti-poisoning security posture:

• Anti-poisoning memory: memories extracted automatically are pending by default — they only enter recall after the user explicitly confirms. Malicious or erroneous content cannot silently pollute later sessions.
• LLM same-origin principle: the configured apiKey determines the trust source; the engine never mixes keys across providers, preventing key hijacking.
• Local-first data: memory lives in ~/.proma-proactive/ as plain local files (JSONL/markdown). No cloud sync, no server.
• Zero outbound by default: memory_capture, memory_recall, suggest_* are pure local rules and never call out. memory_extract's rule mode is zero outbound; LLM mode only sends the current conversation snippet to the LLM endpoint you configured (DeepSeek-compatible by default).
• Secrets never echoed: LLM config is read from .env / environment variables and is never printed to logs or returned by tools.

Please report vulnerabilities privately via GitHub Security Advisory:

https://github.com/ConradLu2740/ProactiveAgent/security/advisories

Do not open a public issue for security problems.

We will acknowledge receipt within 48 hours, and work on a fix as soon as possible. You may be credited in the advisory if you wish.