Tool-Invocation Policy and Autonomous Guardrail Brief

Role: Lead AI Safety and Agent Governance Editor specializing in autonomous system containment, security policy, and tool verification gates.

Context

• Tool Catalog: {{tool_catalog}}
• High-Risk Action Definitions: {{high_risk_actions}}
• Validation Gateways: {{validation_checkpoints}}
• Security & Compliance Standard: {{compliance_standard}}
• Containment & Sandboxing Rules: {{containment_protocols}}
• Deployment Infrastructure: {{runtime_platform}}

Task

Author a comprehensive governance brief defining the tool-invocation policy, runtime validation guardrails, and execution boundaries that restrict autonomous agents from taking unverified or harmful actions.

Method

1. Categorize all entries in {{tool_catalog}} into tiered risk bands (read-only, idempotent write, destructive action).
2. Isolate all operations matching {{high_risk_actions}} and formulate required pre-execution authorization challenges.
3. Integrate {{compliance_standard}} requirements into explicit verification filters applied to raw LLM tool arguments.
4. Define interceptor logic for {{validation_checkpoints}} to scrub, sanitize, and validate tool payloads prior to dispatch.
5. Formulate side-effect containment policies using {{containment_protocols}} for sandboxed execution in {{runtime_platform}}.
6. Detail prompt-injection defense layers to prevent indirect attacks from subverting tool call parameters.
7. Establish strict audit logging requirements, telemetry capture, and breach alerting criteria.
8. Produce clear policy statements directly translatable into programmatic middleware assertions.

Constraints

• MUST mandate human confirmation or multi-party verification for every destructive write action in {{high_risk_actions}}.
• MUST NOT allow silent tool execution bypasses under any operational mode.
• MUST align all audit telemetry specifications with {{compliance_standard}}.
• Language must be authoritative, unambiguous, and directly actionable by software security teams.
• Document length must remain between 850 and 1,250 words.

Output format

Produce a structured policy brief formatted as:

1. Policy Scope and Tool Classification Matrix
2. Mandatory Pre-Execution Validation Gates (Argument Scrubbing, Signature Checks)
3. High-Risk Action Safeguards & Authorization Matrix
4. Containment, Sandboxing, and Blast-Radius Limits
5. Telemetry, Audit Trails, and Incident Response Triggers

Self-review

• Are all destructive capabilities in {{tool_catalog}} explicitly categorized with gating criteria?
• Does the verification logic account for indirect prompt injection vectors via retrieved context?
• Are audit logging requirements sufficient to reconstruct agent decision chains post-incident?