clavenar-agent-sdk (Python) is the client half of a security product. We take vulnerability reports seriously and aim to acknowledge every report within 72 hours.

Email vanteguardlabs@gmail.com with:

• A description of the issue and the impact you observed.
• Steps to reproduce. A minimal proof-of-concept is appreciated but not required if the issue is structural.
• Affected file path, commit hash, and (if applicable) the ClavenarOptions configuration that reproduces the issue.
• Whether you would like public credit in the disclosure announcement.

PGP/GPG: not yet available. If you need an encrypted channel, mention it in your initial email and we will arrange one.

In scope:

• The clavenar_agent_sdk package: client-side request shaping, the wrap surface (clavenar_wrap), retry / pending poll loops, and the ClavenarDenied / ClavenarPending / ClavenarTransportError raise contract.
• Transport security between the SDK and the inspect endpoint (endpoint URL handling, TLS verification posture, header forwarding via extra_headers).
• The streaming intercept (Anthropic content_block_stop / OpenAI finish_reason="tool_calls") — verdict-before-tool ordering.
• Sync vs async client detection in clavenar_wrap.

Out of scope:

• Authentication against upstream model providers (Anthropic API key, OpenAI API key). Those flow through the upstream SDKs unchanged.
• Sandboxing the Python runtime itself. The SDK runs in your process; arbitrary Python is out of our trust boundary by construction.
• Issues in httpx, anthropic, or openai upstream — please report to those projects directly. We track CVE advisories that affect pinned versions in CHANGELOG.md.
• Findings against the demo flow on demo.clavenar.com when caused by demo-specific configuration (the demo accepts X-Clavenar-Demo-Prefix headers visitors mint themselves).

We will not pursue civil or criminal action against researchers who:

• Make a good-faith effort to avoid privacy violations, destruction of data, and interruption or degradation of our services.
• Only interact with accounts they own or with explicit permission of the account holder.
• Give us reasonable time to respond before disclosing publicly.
• Do not exploit a security issue beyond what is necessary to confirm it.

• 72 hours: acknowledgement of the report.
• 7 days: triage outcome (accepted / duplicate / out-of-scope) and a CVE assignment plan if applicable.
• 90 days: public disclosure, coordinated with the reporter.

We may extend the disclosure window for issues that require a coordinated multi-language fix (the TypeScript SDK at clavenar-typescript-sdk shares the same wire contract); we will tell you in advance and explain why.