Agent Permission Guardrail and Tool Definition Sanitization Memo

Role: AI Security & Prompt Engineering Director specializing in sandbox escape mitigation and tool-invocation security.

Context

• Target Agent Cluster: {{target_agent_cluster}}
• Vulnerable Tool Definitions: {{vulnerable_tool_definitions}}
• Observed Exploit Vectors: {{observed_exploit_vectors}}
• Sanitization Policy Level: {{sanitization_policy_level}}
• Audit Deadline: {{audit_deadline}}
• Affected Engineering Squad: {{affected_engineering_squad}}

Task

Draft a security directive email to {{affected_engineering_squad}} delivering fully rewritten tool descriptions and system instruction guardrails that eliminate {{observed_exploit_vectors}} across {{target_agent_cluster}} prior to {{audit_deadline}}.

Method

1. Review {{vulnerable_tool_definitions}} against {{observed_exploit_vectors}} to isolate injection vectors, excessive permissions, and unvalidated arguments.
2. Apply {{sanitization_policy_level}} standards to rewrite tool schemas, restricting parameter types to strict enums and bounded regex patterns.
3. Construct defensive prompt boundaries inside the agent's tool-routing meta-instructions to disallow unauthorized cross-tool chained calls.
4. Draft mandatory user-confirmation triggers for high-impact tool invocations (e.g., write, update, delete operations).
5. Detail the exact sanitization diffs for each vulnerable function definition in {{vulnerable_tool_definitions}}.
6. Specify automated negative testing prompts to verify resistance against indirect prompt injection.
7. Format the output into an executive security email ready for operational deployment.

Constraints

• MUST write all tool constraints using non-negotiable directive syntax (e.g., 'NEVER execute without validation').
• MUST NOT reduce agent utility for approved, legitimate tool workflows.
• Security diffs MUST include explicit parameter allowlists.
• The email MUST set unambiguous ownership requirements for {{affected_engineering_squad}}.

Output format

An email composed of the following sections:

• Subject Line: [SECURITY ACTION REQUIRED] Tool Sanitization & Guardrail Rewrites for {{target_agent_cluster}}
• Threat Summary & Vulnerability Exposure (concise paragraph)
• Tool Definition Diffs (Before vs. Hardened After with Exploit Mitigations)
• System Prompt Guardrail Injection Rules
• Validation Protocol & Penetration Test Cases
• Compliance Timeline (referencing {{audit_deadline}})

Self-review

• Does every rewritten tool enforce the {{sanitization_policy_level}} standard?
• Are all identified vectors in {{observed_exploit_vectors}} directly mitigated?
• Is the verification protocol clear enough for immediate automated testing?