AI safety and auditability
Not every capability described on this wiki is available today.
Quick answer
In Mitra OS, language-model calls can only read and draft. They work from live data, every call is audit-logged, and a failure in one domain does not spread.
Summary
The platform limits what AI can do by design. Models read information and draft responses; actions that matter wait for your confirmation. Grounding in live data reduces invented answers, and audit logs record what was asked and done.
Key takeaways
- Language-model calls are read and draft only.
- Answers are grounded in live data.
- Calls are audit-logged.
- Domain failures are isolated.
Design rules
- Read and draft only for language models.
- Ground answers in live data.
- Log calls for audit.
- Isolate domain failures so one problem stays contained.
Why it matters
An assistant that can both decide and act is risky. Separating drafting from acting keeps you in charge. See Step-up confirmation and AI agents.
Frequently asked questions
People also ask
- What is an AI agent in Mitra OS? (AI agents in Mitra OS)
- Which domains do agents cover? (AI agents in Mitra OS)
- Can an agent move my money on its own? (AI agents in Mitra OS)
- When does Mitra ask me to confirm? (Step-up confirmation)
- Is being signed in enough? (Step-up confirmation)
Related topics
feature AI agents in Mitra OS How Mitra OS uses an orchestrator and specialist AI agents to turn one request into one reconciled answer, and which actions need your confirmation. feature Step-up confirmation Why Mitra OS asks for a fresh biometric confirmation before money, health data or irreversible bookings, even if you are already signed in. guide Security overview An overview of how Mitra OS is designed to protect members: passwordless sign-in, confirmation for sensitive actions, grounded AI and audit logs. feature Passkeys Mitra OS uses passkeys for passwordless, device-bound sign-in with the Secure Enclave on iOS, Android Keystore on Android and WebAuthn on the web.