---
name: agent-output-redaction-gate
description: Use this before publishing, emailing, posting, or handing off agent-generated output when you need to detect and redact secrets, secret values, personal data, private-family references, prompt-injection instructions, or unsafe external-disclosure content.
---
Agent Output Redaction Gate
Run this skill whenever an agent output is about to leave a private workspace: public website copy, social posts, release notes, support replies, customer handoffs, or marketplace listings.
Workflow
1. Save the candidate output to a local text/JSON/Markdown file.
2. Run `python3 scripts/output_redaction_gate.py <file> --format json`.
3. If `decision` is `BLOCK`, do not publish. Remove or rewrite the flagged content.
4. If `decision` is `REVIEW`, ask a human owner before publication.
5. If `decision` is `ALLOW`, use the original text or the generated `redacted_text` if you want conservative masking.
What it checks
→ API keys, bearer tokens, private keys, provider key formats, passwords, and environment-file leaks.
→ Personal contact details and private-family/publication boundary markers.
→ Prompt-injection text that tries to override system/developer instructions.
→ Financial-action language that could imply unsupervised transfers, trades, or payments.
→ Sensitive internal configuration or memory files that should not be quoted publicly.
Safe use
The script is local-only. It performs no network calls, reads only the file you pass, and never mutates the source file unless you explicitly write the redacted output to a separate path.
Release evidence
Reviewed in the 2026-06-15 GetAgentIQ warehouse propagation cycle.