Only the latest released version of Clawness is supported. Since marketplace.json points at main (see CLAUDE.md), updating to the newest release is the only supported way to get a fix.

Please report security issues privately via GitHub Security Advisories rather than a public issue. Include:

• The affected file(s)/hook(s) and version
• Steps to reproduce, and what you expected vs. observed
• Impact (what an attacker could do with it)

You should get an initial response within a few days.

Clawness runs entirely locally as a Claude Code plugin — there's no server or service to attack. Relevant classes of report:

• A hook (hooks/*.py) that can be tricked into executing, reading, or transmitting something it shouldn't
• The access guard (clawness/guard.py) failing to catch (or wrongly allowing) a genuinely dangerous tool call
• The trust ledger (clawness/trust.py) missing a real prompt-injection pattern in a skill/agent/MCP artifact
• Path traversal or arbitrary file write via rule/memory/handoff file handling

Out of scope / by design: the access guard is a harm-reduction tripwire over an agent's own tool calls, not a sandbox — it's built to catch honest mistakes and low-effort attacks, and a determined adversary who already controls the agent's shell can route around regex-based detection (see CLAUDE.md for the full threat-model rationale). If you find a guard bypass, it's still worth reporting — we track and close specific bypasses even though the guard doesn't claim to be airtight — just note in the report whether it's a design gap or a genuine bug in the matching logic.