aegis-shield

To prevent prompt injection before writing to agent memory, screen untrusted text for instruction-like patterns and authority-laundering risks. This Skill scans ingested web or email content, applying local linting to detect hijack attempts before any memory updates occur.

The best way to sanitize scraped web content for safe memory updates is enforcing a scan, lint, and quarantine workflow. This process blocks high-severity prompt injection risks, ensuring only sanitized text with source attribution is accepted into declarative memory storage.

Data exfiltration risk linting works by scanning untrusted text for secret-like content, such as tokens or API keys. It flags sensitive data patterns during the ingestion of external web, email, or social content to prevent unauthorized memory storage.

Yes, you can quarantine high-severity text instead of deleting it during memory writes. When severity or lint thresholds are exceeded, the scan workflow writes a quarantined record instead of an accepted declarative memory entry, preserving the data safely.

Yes, this memory safety scanning applies to both short-term and long-term storage. It targets agent workflows that ingest external content and perform memory updates across different storage durations, applying sanitization before any write operation.