LG-SOTF processes security telemetry and can trigger containment actions, so we treat vulnerabilities in it seriously. Thank you for reporting responsibly.

Release channels (weekly partner candidates, stable, LTS), their support windows and which fixes are backported: docs/releases.md.

Do not open a public issue, pull request or discussion for a security problem.

Report it privately through GitHub's private vulnerability reporting: open the repository's Security tab and choose Report a vulnerability (https://github.com/gapilongo/SOC/security/advisories/new). If that option is not available to you, open an issue that only asks a maintainer to contact you privately, without any details of the problem.

Please include:

• the affected version or commit, and the deployment mode (Compose, Helm, other);
• the component (API, worker, scheduler, migrations, Helm chart, web app, evaluation harness) and the configuration needed to reproduce;
• steps to reproduce or a proof of concept, and the impact you expect (for example cross-tenant data access, authentication bypass, unintended response action, secret disclosure, prompt injection that closes a malicious alert or triggers containment);
• whether the issue is already public or known to others.

Use test data only. Do not access data that is not yours, do not run response actions against systems you do not own, and do not degrade shared services while testing.

• We acknowledge a report within 5 business days and give a first assessment within 10 business days.
• We keep you informed while we work on a fix and agree a disclosure date with you; we aim to release fixes for critical and high severity issues within 30 days of confirmation.
• We publish a GitHub security advisory with the fix and credit you if you wish.

In scope: the code in this repository, the container image built from docker/Dockerfile, the Helm chart in deploy/helm/lg-sotf, and the documented default configuration.

Particularly relevant: tenant isolation (row-level security, Redis keys, checkpoints, WebSocket events), authentication and session handling (JWT, OIDC, API keys, refresh token rotation), authorisation of escalation decisions and action approvals, response guardrails (protected assets, autonomy policy, rate limits), secret handling and redaction, the outbound destination policy of pull connectors (SSRF), webhook signing, audit log integrity, and prompt injection that changes a triage outcome.

Out of scope: vulnerabilities in third-party services (LLM providers, SIEM and EDR products, SOAR platforms) and in dependencies that are not exploitable through LG-SOTF (report those upstream); findings that require an already compromised host, database or Kubernetes cluster; missing hardening on deployments that ignore the production safety checks (for example environment=development exposed to the internet); denial of service by volume alone; the deterministic fake model's classifications (it is a test double, refused in production).

The threat model, tenant isolation, secrets handling, prompt-injection defences, response guardrails, data handling and audit are described in docs/security.md.