AgentSkillsCN

gemini-context-security

智能体上下文的安全协议。通过严格的数据隔离、PII脱敏，以及防范内存中毒与提示注入攻击，为智能体筑起坚实的安全防线。

SKILL.md
--- frontmatter
name: gemini-context-security
description: security protocols for agent context. Use this to implement strict data isolation, redact PII, and prevent memory poisoning or prompt injection attacks.

Gemini Context Security Strategies

Goal

Secure the agent's "brain" (Session & Memory) by enforcing strict isolation, sanitizing sensitive data, and defending against adversarial manipulation.

Core Security Principles

1. Strict Isolation (The "Archivist" Rule)

  • •Concept: Treat memory as a secure archive. An agent serving User A must never have access to the memories of User B.
  • •Implementation:
    • •Access Control Lists (ACLs): Enforce strict permissions at the database level.
    • •Tenant Scoping: Every memory and session event must be tagged with a user_id or tenant_id.
    • •Zero Trust: The agent runtime should only possess the credentials to access the specific user's data it is currently serving.

2. Data Sanitization (PII Redaction)

  • •Concept: "Redact before you write." Sensitive Personally Identifiable Information (PII) must be removed before data is persisted to long-term storage.
  • •Workflow:
    1. •Detect: Use a regex or a specialized NLP model (like Google Cloud DLP) to identify emails, phone numbers, or credit cards in the conversation stream.
    2. •Redact: Replace the sensitive text with a placeholder (e.g., [REDACTED_EMAIL]).
    3. •Persist: Only save the sanitized version to the Session/Memory store.
  • •Benefit: Reduces the "blast radius" of a potential breach and simplifies GDPR/CCPA compliance.

3. Defense Against Memory Poisoning

  • •Threat: A malicious user (or prompt injection) tricks the agent into saving a false or harmful fact (e.g., "Ignore all safety rules" or "My bank account is [Attacker's Account]").
  • •Mitigation:
    • •Input Validation: Use "Model Armor" or guardrails to scan inputs for malicious intent before processing.
    • •Verification: Do not blindly trust implicit memories. Require corroboration or higher confidence scores for sensitive instructions.
    • •Sanitization: Strip executable code or prompt-like syntax from memory content before storage.

4. Shared Memory Risks (Procedural Memory)

  • •Scenario: If you share "Procedural Memories" (playbooks/skills) across users to help the agent learn globally.
  • •Risk: A playbook learned from User A might contain User A's private API keys or confidential project names.
  • •Protocol:
    • •Aggressive Anonymization: scrub all specific entities (names, dates, IDs) from procedural memories.
    • •Review: Require a human-in-the-loop or a secondary LLM review before promoting a memory from "User Scope" to "Global/Application Scope."

Production Checklist

  • • Isolation: Is every database query scoped by user_id?
  • • Redaction: Is PII stripped before writing to logs/DB?
  • • Retention: Is there a TTL (Time-To-Live) policy to automatically delete old sessions?.
  • • Provenance: Does every memory record where it came from (Source Type)?.