agent-contracts is a deterministic guardrail library for agent tool calls and outgoing responses. It is not a sandbox. Security reports should focus on cases where the package claims to block or warn but fails under a reasonable sanitized input.

The project is pre-1.0. Security fixes target main first. Release tags will define supported versions once the package is published.

Useful security reports include:

• Secret patterns that should be caught but are missed.
• Dangerous paths that the default path guard should block but allows.
• Inputs that cause a contract to crash instead of returning a violation/pass.
• Packaging issues that ship different code from the repository.
• Documentation that materially overstates the security boundary.

Not security issues for this package:

• A tool performs unsafe behavior after Registry.enforce_pre() allowed it.
• A host application ignores BlockedAction.
• An agent sees secrets before the contract receives an ActionContext.
• A custom contract written by an application author is incomplete.

Those belong in the host application's sandboxing, authorization, credential scoping, or custom policy layer.

Use a private channel if the issue includes a live credential, exploit path, or non-public system detail. Otherwise, open a GitHub issue with a sanitized minimal ActionContext, expected behavior, and actual behavior.

Never paste real secrets into an issue. Replace them with same-shape test values.

Security fixes should include:

1. A regression test that fails before the fix.
2. A narrow contract or pattern change.
3. Documentation updates if the threat model or recovery guidance changed.