What problem does it solve? Agents that read cloned repositories, web search results, or fetched files can be hijacked by prompt-injection text embedded in that content. This Skill screens untrusted content before any agent consumes it, neutralizing injection attempts and preventing malicious instructions from reaching the model. ## Core Features & Use Cases - Hybrid detection pipeline: Combines a six-category regex pattern catalog (role-hijack tokens, tool-use directives, exfiltration, hidden content, encoded payloads) with an LLM analysis pass that confirms hits, catches false positives, and extends spans to cover full malicious directives. - Severity-keyed actions: Low and medium severity passages are replaced with safe [REMOVED: <pattern_class>] markers; any high-severity hit aborts sanitization entirely so the content is never consumed. - Persistent flagged-sources tracking: Maintains docs/security/flagged-sources.md as a git-tracked audit trail, automatically bumping caution levels for repeat-offender files, repos, and domains. - Use Case: Before synthesizing research from a cloned third-party repo, run the sanitizer on each file; injection attempts are stripped or the file is quarantined, and the source is logged for future runs. ## Quick Start Sanitize the content fetched from this external URL before summarizing it, and report any injection attempts found.
