stop-that-shit

Least-privilege permissions

• Limit what the agent can access: Only grant the "stop-that-shit" tool permission to access the specific files and systems it actually needs—not your entire codebase or infrastructure.
• Restrict filter rule editing: Lock down who can create or modify the blocking rules; treat these like security policies, not developer conveniences.
• Test rules in isolation first: Run new filter configurations in a sandbox or low-stakes environment before deploying them to production systems.
• Audit filter exceptions regularly: Review any requests that bypass the guards monthly; document why exceptions were granted and whether they're still needed.

Network isolation

• Keep the agent sandboxed: Run the AI coding agent in a contained environment (isolated container or virtual machine) so if filtering fails, it can't reach sensitive databases or external systems without additional authorization steps.
• Monitor outbound connections: Log all network calls the agent makes; catch unexpected attempts to contact external services or pull undocumented dependencies.
• Require manual approval for new packages: Don't let the agent automatically fetch libraries or tools; require a human security review first, especially if the tool's dependency list isn't well-known.

Monitoring

• Log every filter decision: Record what the agent tried to do, what the filter blocked or allowed, and when—create an audit trail even if the tool doesn't currently provide detailed transparency.
• Alert on unusual patterns: Set up warnings if the agent suddenly requests many hashes, checksums, or system information you didn't ask for; this could indicate prompt injection or unexpected behavior.
• Track false negatives: Keep records of harmful outputs that weren't caught; use this feedback to improve or reconfigure the guard rules over time.
• Review vendor updates: Check the tool's GitHub or release notes weekly; security gaps or rule changes need to be understood and tested before they reach your systems.

Red flags to watch for

• Silent failures: If the agent stops responding without clear error messages, the filter may be blocking legitimate work; investigate rather than assume everything is fine.
• Rule opacity: If you can't read or understand why a request was blocked, escalate to the vendor for documentation; unexplained filtering is a liability.
• Scope creep in blocking: If the tool starts filtering requests that seem unrelated to "hashes and checksums," question whether the rules are drifting beyond their intended purpose.
• No clear rollback plan: Before deploying, ensure you can quickly disable the tool or revert rule changes if it causes problems; missing rollback procedures are a major risk.