Security reports are accepted for the current main branch of github.com/muonsoft/llm-guard while the project is in pre-1.0 MVP stabilization.

In scope:

• PII or secret leakage through public APIs, observers, errors, or logging defaults
• Bypass of default secret blocking or policy enforcement
• Mask/restore correctness issues that expose caller-owned sensitive values
• Supply-chain disclosure gaps for shipped dependencies or project-authored data

Out of scope for this repository:

• Third-party LLM provider security
• Deployment/network perimeter controls outside the library
• Issues requiring real PII, live credentials, or copied upstream dictionaries to reproduce

Do not open a public GitHub issue for security vulnerabilities.

Report privately to the maintainers through your organization's agreed secure channel for MuonSoft open-source projects. If you do not have a contact path, open a minimal private security advisory request through GitHub Security Advisories for muonsoft/llm-guard and omit sensitive payload values.

Include:

• affected version or commit
• reproduction steps with synthetic fixtures only
• impact assessment
• suggested fix if available

Allow reasonable time for triage and remediation before public disclosure.

• Never include real PII, credentials, private keys, or production text in public issues, pull requests, comments, or attached logs.
• Use synthetic tokens and fictional names/addresses from project test patterns.
• Do not paste TokenSet contents or masked mappings into public channels.

• Default observer is a no-op; safe observers expose bounded metadata only.
• Secrets block Mask by default (ErrBlocked); masking requires explicit opt-in.
• WithUnsafeDevelopmentObserver is for local debugging only and must not ship in production paths.

Maintainers may request additional evidence under the same synthetic-data rule.