Security
How NeverHalt protects your agents, keys and data, and what is not covered yet.
This page lists what is implemented and tested. It is also honest about what is not.
What we do
- Secrets and AI provider keys are encrypted at rest (AES-256-GCM). Provider keys are never returned by the API and never enter your agent container.
- Every agent has its own token. It can send events, call the gateway and ask for approvals for that agent only. Tokens can be rotated.
- The AI gateway only forwards allow-listed endpoints, adds your key after the checks, and enforces the daily budget.
- Only the signed-in owner can approve an agent request. API keys and assistants cannot.
- Agents run in containers with dropped capabilities, no privileged mode, memory and process limits, and no access to cloud metadata addresses.
- Passwords are hashed with PBKDF2; sessions are stored hashed; sign-in, registration and password reset are rate limited and protected against automated abuse.
- We do not store the text of your prompts or the model answers.
What is not covered
- There has been no independent penetration test yet.
- The token of an agent is a secret known to the agent code itself: whoever controls the code can use it, within that agent budget and events only.
- Spend limits only apply to calls that go through the gateway.
Report a problem
Write to security@neverhalt.app. See /.well-known/security.txt.