|
NeuralPlus 0.2.0
A small C++17 SDK for AI clients, sessions, tools, and tracing
|
NeuralPlus is pre-alpha. Security fixes are applied to the latest development branch until versioned release support begins.
Do not open a public issue for a vulnerability, leaked credential, unsafe default, or trace containing private content. Use GitHub private vulnerability reporting for this repository. If it is unavailable, contact the repository owner privately through the contact method on the GitHub profile.
Include the affected commit, a minimal reproduction, impact, whether credentials or personal data are involved, and a proposed fix when available. Do not attach working credentials or unsanitized production traces.
ProviderError fields. Common credential header names are recognized automatically; custom credential headers must set HttpHeader::sensitive.capture_trace_payloads. In-memory, callback, and custom tracers then receive the payload directly. FileTracer requires its separate include_payloads option; console and syslog remain metadata-only.0600; supported POSIX platforms also reject final-component symbolic-link destinations.Payload capture can expose prompts, model output, tool arguments, personal data, or credentials returned by a tool. If enabled, apply access controls, retention limits, encryption, and redaction appropriate to the application. FileTracer::max_bytes is a size guard, not rotation or retention management. Exact credential redaction cannot recognize transformed values. An injected transport is responsible for keeping request credentials out of its own logs and non-ProviderError exceptions. MockHttpTransport intentionally records complete requests for tests, including authentication headers, so use only synthetic credentials and sanitized fixtures.
The transport and generation limits are resource guards, not full request quotas, rate limiting, cancellation, or application-level authorization.
The project has not received an independent security audit. Consumers remain responsible for threat modeling and review before sensitive deployment.