aegis-shield
от deegerwalker
Web & Frontend DevelopmentКратко
Prompt-injection and data-exfiltration screening for untrusted text.
Описание
Что делает этот навык
Screens untrusted text for prompt injection, data exfiltration, and tool-abuse patterns before it reaches memory or other tools. Assigns a severity score and routes content to either acceptance or quarantine. Includes a safe memory append workflow that enforces scan and lint checks before any write.
Intercepts malicious content at the boundary rather than relying on downstream tools to handle hostile input safely.
Когда его использовать
- Scanning email content before summarizing it to an agent
- Screening scraped web pages before feeding them into a pipeline
- Quarantining social media replies that contain suspicious instructions
- Enforcing sanitized memory writes when processing user-supplied input
- Blocking exfiltration attempts hidden in LLM-targeted document payloads
Пример сценария
User asks: scan this webpage content before summarizing it
- 1
Pass the raw fetched text to the scan step
- 2
Receive severity score and flag list from the scanner
- 3
Check if severity is medium or higher
- 4
If severity is medium+, route content to quarantine and halt pipeline
- 5
If severity is low, proceed to summarize the sanitized text
Content is either quarantined with a report or cleared for downstream use
Установка
clawhub install deegerwalker/aegis-shieldВыполните команду CLI или вставьте промпт в Claude Code, Cursor, Codex — любой кодинг-агент, которым пользуетесь.