Português

This repository distributes the security-sensitive parts of a Claude Code setup: hooks that block destructive commands and sensitive-file access, secret-shaped-content scanning, MCP governance guardrails, and a Bandit/pip-audit-backed quality gate for generated projects. Treat vulnerabilities in this repository - not just in the applications it scaffolds - as security issues worth reporting responsibly.

• Bypasses of the PreToolUse/PostToolUse hooks under template/.claude/hooks/ and plugin/python-engineering-harness/scripts/ (protect_sensitive_files.py, validate_bash.py, guard_mcp.py, scan_secrets.py) - e.g. a destructive command pattern, a sensitive path, or a secret shape that should be blocked but isn't.
• Logic errors in template/scripts/validate_architecture.py or template/scripts/validate_mcp_config.py that let a violating configuration or dependency pass the gate.
• bootstrap.py behavior that could write outside the intended target directory, clobber files under --merge in a way the documented .harness-new conflict path does not describe, or otherwise defeat the "never overwrite existing files" guarantee.
• MCP governance gaps: anything that would let a configured server exfiltrate secrets past guard_mcp.py, or that would let a state-mutating tool call skip the confirmation requirement documented in template/docs/MCP.md.
• Any documented control in docs/ENTERPRISE_ROLLOUT.md (allowlist/denylist enforcement, managed settings examples) that does not behave as described.

• Vulnerabilities in third-party dependencies pulled into a bootstrapped project (fastapi, sqlalchemy, etc.) - report those upstream; pip-audit in the generated project's quality gate is the intended detection mechanism, not this policy.
• Findings that require the operator to have already disabled a documented safety control (for example, an explicitly-approved bypassPermissions mode, or a hand-edited hook that removes a check) - the harness's threat model assumes its shipped controls are left in place.
• Issues in Claude Code itself, as opposed to this harness's configuration of it - report those to Anthropic through the official Claude Code channels.

This project does not yet publish a dedicated security contact address. Until one is added here:

• If this repository is hosted on GitHub, use GitHub's private vulnerability reporting (Security tab -> Report a vulnerability) so the report is not publicly visible before a fix ships.
• Otherwise, contact the repository owner listed under "Recommended ownership" in docs/ENTERPRISE_ROLLOUT.md (platform engineering for hooks and CI baseline, security for blocked paths and secret patterns) through your organization's internal channel.

Please do not open a public issue for a suspected vulnerability before the maintainers have had a chance to assess and, where warranted, ship a fix.

• The specific hook, script, or bootstrap flag involved, and the exact input or command that should have been blocked and was not (or that produced unexpected write behavior).
• Whether the issue reproduces against template/, plugin/python-engineering-harness/, or both.
• Your Claude Code version and OS, since hook matching and shell behavior can differ across platforms (template/README.md and this repo's README.md both note macOS/Linux/WSL as the supported set for the default hook configuration).

Given this is a template/plugin distribution rather than a hosted service, there is no user data to protect directly, but a missed hook bypass or gate failure could propagate into every project that bootstraps from an affected version. Please allow time for a fix and a CHANGELOG.md entry (see the Security section convention already used there) before public disclosure.