English | 日本語
Use GitHub's private vulnerability reporting: the repository's Security tab → Report a vulnerability. Please do not open a public issue, discussion or pull request for a vulnerability. Include what you can of:
- the version or commit you tested;
- steps to reproduce (a minimal configuration and the prompts or requests used);
- the impact: which boundary is bypassed and what an attacker gains;
- the environment (OS, Python version, model provider, chat server settings that matter).
Do not paste real secrets, keys or personal data; use obvious dummies (sk-test-…).
We ask for coordinated disclosure: please wait for a fix or an agreed date before publishing.
StiraCore is pre-alpha. Only the latest main is supported; fixes are not backported.
Handling is best effort, with no response or fix-time commitment.
A way to bypass a boundary the harness says it enforces, for example: secret masking,
the secret-path floor and protected_dirs, IdentityGuard, tool consent and approvals
(Always / Ask / Never, approval binding), the YOLO window limits, the circuit breaker,
the SSRF / outbound-request restrictions, and the local web UI's Host / Origin / CSRF
checks and API permissions. See the security model,
tool execution paths,
API permissions and containment.
- The harness is not an OS sandbox (the security model lists the measured limits; containment phase 0 observes only).
- An attacker who already runs arbitrary code as the same OS user, or a compromised OS.
- The quality or truthfulness of model output by itself, without a boundary bypass.