mirror of
https://github.com/openclaw/openclaw.git
synced 2026-10-26 10:13:27 +08:00
cefcfad4b8
* fix(security): audit host exec the way the runtime resolves it (#183541) The audit resolved an unset tools.exec.security as "deny", while the runtime (prepareExecDefaults) resolves it as "full" for an unsandboxed agent. An agent with no exec settings, or only ask: "off", ran host exec with full access and no approvals while `openclaw security audit` reported nothing, and security.exposure.open_channels_with_exec left it out of the reachable scopes and stayed at warn. Resolve every scope through resolveExecDefaults, as the skill/MCP boundary check already does. The exposure check now counts and rates every reachable scope whose effective security is full, including the unsandboxed default. tools.exec.security_full_configured fires when that full trust is an operator choice: an explicit full mode/security, or approvals switched off; a bare default install without open channels stays clean. * fix(security): audit only configured exec agents --------- Co-authored-by: Peter Steinberger <steipete@gmail.com> Co-authored-by: lonexreb <reach2shubhankar@gmail.com>