# AI Security & Safety Notice for Members **Effective date:** 2026-08-03 This notice explains how the Quantum Aikido coaching chatbot keeps your conversations safe, what the AI can and cannot do, and where a human coach takes over. It is designed to be read alongside the [Corpus Privacy Policy](corpus-privacy-policy.md). --- ## 1. What the chatbot is The coaching chatbot is an AI assistant grounded in a private, member-only corpus of teaching material. It is not a general-purpose search engine or open chat service. Access is restricted to registered members who have received an invitation code. The chatbot answers in the voice of the teachings. It does not attribute responses to any named person and will not quote a teacher or participant by name. ## 2. Your messages are private and used only to help you - Your chat queries are stored so you can resume conversations and so we can validate system updates. - They are shared only with Google Gemini, the AI provider used to generate your response. No other third party receives your chat text. - You may request deletion of your conversation history at any time by contacting **kenneth@quantumaikido.com**. We aim to complete deletion within 14 days. See the [Corpus Privacy Policy](corpus-privacy-policy.md) for the full details on data retention, exclusion rights, and redaction. ## 3. Protections against misuse of the AI We use automated checks to: - **Block attempts to override the coach's instructions.** Messages that look like "ignore previous instructions" or "you are now a different AI" are detected and routed to a human coach instead of the model. - **Strip hidden manipulation.** Invisible characters, encoded text, and fake structure markers are removed before your message reaches the AI. - **Keep responses grounded.** The AI only answers from the private corpus. If a question has no relevant grounding material, the chatbot will offer to connect you with a human coach rather than invent an answer. These filters are conservative. Occasionally a normal coaching question may be misclassified as manipulation and handed off to a human. If this happens, you can rephrase the question or wait for a coach reply. ## 4. When a human coach is brought in The chatbot hands off to a human coach automatically when: - You ask for a human coach. - The question involves sensitive personal matters that are better handled person to person. - The question falls outside the private teaching material. - An automated safety filter is triggered. A handoff creates a coaching room link. Do not share that link with anyone outside the community. ## 5. Known limitations we are working on The system is actively monitored and improved. Current areas of work include: - **Conversation integrity.** Client-side chat history is currently used as a hint for context. We are adding server-side integrity checks so the model only sees validated history. - **Cross-instance rate limiting.** Rate limits currently protect a single running instance. We are moving them to a shared store so protection remains consistent when multiple instances are running. - **Session authorization under outages.** A database error during session checks must not grant access. We are hardening this path so the system fails closed. - **Link validation.** "Join call" links and video embeds are being reviewed to prevent malicious URL tricks. - **Admin dashboard rendering.** Admin tables that display corpus filenames are being updated to prevent browser-script injection. - **Credential hygiene.** A previously exposed YouTube API key is being rotated; it does not affect member login or chat data. ## 6. How to report a security or safety concern If you see the chatbot behave unexpectedly, leak internal terminology, reveal someone's words without consent, or produce harmful content, please report it to **kenneth@quantumaikido.com**. Include: - The date and approximate time. - What you typed (or a description if you prefer not to repeat it). - What the chatbot did. We review every report and use it to improve filters, prompts, and access controls. ## 7. Changes to this notice Material changes will be logged here with a date. Minor clarifications may be made without a new date. - **2026-08-03** — Initial AI security and safety notice, summarizing prompt-injection filtering, handoff behavior, session protections, and active security improvements.