alinafe82 Agent Security Boundary
Prevent agent workflows from widening permissions, leaking data, or trusting hostile context.
Scope
Use for the task in the description. For example: Review a tool that executes commands extracted from uploaded documents. Do not activate for: Rename a local helper without changing tool or data access.
Workflow
- Map every tool, file, network, and external-state boundary.
- Classify data sensitivity and untrusted inputs.
- Identify how prompt injection or misleading context could steer tool use.
- Check approval gates for destructive, permission-widening, or external-send actions.
- Recommend guardrails, rejection, or targeted validation before merge.
Evidence
Tool, file, network, or permission surfaces touched. Data sensitivity and trust boundaries. User approvals or destructive action requirements. Prompt-injection and untrusted-content exposure. Report the outcome with relevant workflow, tools and data, trust boundaries, prompt-injection risks. Adapt the format to the task; omit empty fields. Never claim checks ran without observed results.
Boundaries
Keep secrets and private records out of shared artifacts. Use redacted or synthetic evidence. External sends, publication, destructive operations and permission widening need explicit authorization; reuse authorization already given for the action. Continue authorized read-only and reversible local work through the requested outcome. If evidence is missing, inspect available sources before asking; report limits honestly.
References
For worked examples, open simple or edge case only when useful. Routing cases record positive and negative activation examples for review; they are not a model benchmark.