Guard mode
mcpwn is an offensive tool by default, but the same message plumbing runs the other way too. Point it defensively and it becomes a protective firewall on the agent↔server channel. These three features are read-only or fail-open - they make traffic safer or leave it unchanged, never deadlocking the stream.
Guard mode (--guard / --policy <file>)
Section titled “Guard mode (--guard / --policy <file>)”The defensive inverse of the tamper proxy. For every parseable message mcpwn
recognises the same content classes the passive detectors flag
(prompt-injection phrasing, leaked secrets, exfil URLs, invisible-unicode
smuggling) and, per an active JSON policy (mcpwn.policy.json), decides to:
allow- forward untouchedwarn- forward, but record a findingsanitize- strip only the offending span and forwardblock- drop the message before the agent ever sees it
Guard mode is the one place mcpwn may drop a message - and it fails open: a
buggy policy or pathological payload forwards the original untouched. With no
policy it ships a built-in default that sanitises the four high-severity
classes (least-disruptive: the agent keeps working, only the malicious span is
removed). Precedence: --policy <file> > ./mcpwn.policy.json > built-in
default.
Rug-pull pinning (pin capture + --pin <file>)
Section titled “Rug-pull pinning (pin capture + --pin <file>)”The cross-session cousin of --baseline. mcpwn pin capture hashes a server’s
whole advertised surface - tools and resources and prompts - into a
committed mcpwn.lock; --pin then diffs later runs against it and raises a
HIGH finding on any silent description/schema/metadata change, or any
added/removed item. Because the lock lives on disk, the pin and the check can be
weeks and many processes apart - it catches a server that re-poisons long after
the user approved it. Commit mcpwn.lock to your repo and gate CI on it.
Provenance tainting (--taint)
Section titled “Provenance tainting (--taint)”Makes the indirect-injection kill chain explicit. It tags text originating
from an untrusted server→client tool result / resource as untrusted-external,
then raises a HIGH finding when that same text later reappears inside a
client→server tools/call argument (source message id → sink call). It is
read-only, keeps its own pipeline-owned state, and stays cheap via a bounded,
minimum-length substring index.
# Run as a firewall (built-in default policy: sanitise injection/secrets/exfil/unicode)mcpwn --guard -- node ./my-server.jsmcpwn --policy mcpwn.policy.json -- node ./my-server.js # custom block/sanitise rules
# Commit a lock, then fail later if the advertised surface drifts (cross-session)mcpwn pin capture --out mcpwn.lock -- node ./my-server.jsmcpwn --pin mcpwn.lock --taint --findings-out ./out -- node ./my-server.js
# Pin + taint also work under the self-driving scanner (CI)mcpwn scan --pin mcpwn.lock --findings-out ./out -- node ./my-server.jsFor the defensive framing of each class and its remediation, see the threat taxonomy.