Agentic Harness executes tools, reads and writes project files, stores run evidence, and can connect to user-selected model endpoints. Security reports that could affect workspace containment, credentials, review integrity, the loopback GUI, or the release pipeline are taken seriously.

Security fixes are made against the latest published release and the current main branch. Earlier pre-1.0 releases are not maintained as separate security support lines. Reproduce a report on the latest release when practical, and include the exact version tested.

Do not include exploit details, credentials, private repository content, or a working proof of concept in a public issue.

Use GitHub's enabled private vulnerability-reporting form:

https://github.com/moortekweb-art/agentic-harness/security/advisories/new

If GitHub is temporarily unable to offer that form, open a minimal public issue titled Private security report channel needed. State only that you need a private channel; do not describe the vulnerability. A maintainer can then arrange a private report path before technical details are shared.

Include the following in the private report:

• the affected Agentic Harness version or commit;
• the operating system, Python version, interface, and worker type involved;
• the security boundary that was crossed and the resulting impact;
• minimal reproduction steps using synthetic data;
• whether credentials, external services, or user interaction are required;
• any known mitigations or suggested fix; and
• your preferred disclosure and credit details.

Never send live API keys, access tokens, private model prompts, or proprietary source. Replace them with clearly marked test values.

The project does not promise a fixed response SLA. Maintainers will acknowledge complete reports and provide an initial assessment as soon as practical; complex reports may take longer to reproduce. Please allow a reasonable remediation period and coordinate public disclosure so users can update first.

Confirmed fixes will be tested against the affected boundary and documented in release notes. A GitHub security advisory and CVE may be used when the impact warrants them.

Examples include:

• workspace, symlink, protected-path, or pre-existing-change containment bypasses;
• plaintext credential persistence or credential disclosure through logs, events, reports, URLs, browser storage, or review commands;
• GUI authentication, host validation, origin, or non-loopback exposure flaws;
• acceptance of an unfinished or failed goal as independently verified;
• command, argument, or configuration injection across a documented trust boundary;
• release identity, artifact, trusted-publishing, or immutable-tag bypasses; and
• unsafe file permissions or state corruption that exposes another user's data.

Reports about an explicitly configured external coding agent exercising the permissions granted to that agent are normally outside the embedded engine's security boundary. They are still useful when Agentic Harness misrepresents or bypasses its documented adapter boundary.

Use only systems and data you own or are authorized to test. Avoid privacy violations, persistence, destructive actions, service disruption, and access beyond what is needed to demonstrate the issue. The project welcomes good-faith reports that follow these limits.