The baked-in default policy used when no mcpwn.policy.json is present: deny
(neutralise) the four high-severity content classes by SANITISING them — the
least-disruptive protective action that still removes the threat (the agent
keeps working; only the malicious span is stripped/redacted). Point --policy
at a file to BLOCK instead, or to add JSONPath/tool rules.
The baked-in default policy used when no
mcpwn.policy.jsonis present: deny (neutralise) the four high-severity content classes by SANITISING them — the least-disruptive protective action that still removes the threat (the agent keeps working; only the malicious span is stripped/redacted). Point--policyat a file to BLOCK instead, or to add JSONPath/tool rules.