Security Policy
This file is generated from the canonical repository documentation. Do not edit it directly.
Supported Versions
Akritas is under active development. Security fixes are applied to the current mainline version only until tagged releases and a formal support window exist.
Reporting a Vulnerability
Do not open a public issue for a suspected vulnerability. Send a private report to the repository owner or the security contact configured by the deploying organization. Include:
- the affected revision and deployment mode;
- a minimal reproduction;
- the expected and observed security boundary;
- potential impact and required attacker access;
- any suggested mitigation.
Do not include production credentials, private repository content, or customer data. Operators should define a monitored security contact before exposing an Akritas instance outside a local development environment.
Operational Baseline
- Keep the default loopback bind unless a trusted TLS/SSO reverse proxy protects the service.
- Set
AKRITAS_API_KEYand rotate it through the deployment secret mechanism. - Mount RAG indexes and ordinary workspaces read-only.
- Grant a writable workspace only when Apply is required.
- Treat MCP servers, indexed documents, repositories, Alertmanager payloads, and model output as untrusted.
- Enable
go-testonly inside a separate container or VM security boundary. - Keep write and dangerous MCP tools unavailable to the Web runtime.
- Protect audit storage from modification by the service's ordinary users.
Security Guarantees and Non-Goals
Akritas validates bounded inputs, applies an explicit tool policy, isolates change preparation in a temporary copy, and requires one-time approval before a workspace write. These controls do not turn model-generated code or project tests into safe programs. The temporary workspace and sanitized environment are not an operating-system sandbox.
See Security Model and Threat Model for trust boundaries and residual risks.