# Agent Security Review

> Pre-production security checklist for Agentforce deployments: agent run-as permission scope, grounding-source data classification, action write-scope, audit trail. NOT for the full go-live readiness gate covering cost telemetry, rate limits, rollout and rollback — use agentforce/agentforce-production-readiness-checklist. NOT for a general org security review — use security/security-health-check.

- Skill: `pranavnagrecha/agent-security-review` (Agent Skill, multi-file: 7 files)
- Install (CLI): `npx skillmds add pranavnagrecha/agent-security-review`
- Raw SKILL.md: https://api.skillmd.com/api/skills/pranavnagrecha/agent-security-review/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: PranavNagrecha (https://skillmd.com/u/pranavnagrecha)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/pranavnagrecha/agent-security-review

---


# Agent Security Review

Agentforce agents touch user PII, internal data, and external APIs with the permissions of whatever user invokes them. A structured review covers four axes: (1) least-privilege user, (2) data classification of every grounding source, (3) action write-scope, (4) audit trail completeness.

## Recommended Workflow

1. Export the agent: subagent instructions (subagents were called topics before April 2026), Invocable actions, grounding DMOs/sObjects, channel configs. Checksum and archive.
2. Map the agent's run-as user to a dedicated permission set; verify no profile-level permissions leak through.
3. Classify every grounding source: public, internal, confidential, regulated. Redact/mask confidential+regulated fields at the Trust Layer.
4. Enumerate Invocable write scope: which sObjects/fields can be created/updated/deleted. Apply FLS + CRUD checks in Apex; tighten or split the user.
5. Verify audit trail: every Invocable logs to `Agent_Audit__c`; every conversation is retained per retention policy; Shield Event Monitoring streams cover ApexExecution + ContentTransfer.

## Key Considerations

- Default agent-run-as is the invoking user — this is often over-privileged for admin testers. Use automated user for agent invocation.
- Data Cloud grounding can pull data the invoking user cannot see; verify masking at DMO level.
- Invocables bypass `with sharing` if written as `without sharing` — audit every action class's sharing declaration.
- Audit trail must include prompt + response for forensic replay; conversation storage has retention implications (GDPR).

## Worked Examples (see `references/examples.md`)

- *Write-scope tightening* — Service agent can 'Update any Case field' via a generic UpdateRecord action.
- *Regulated-field masking* — RAG grounding includes Contact.Social_Security_Number__c.

## Common Gotchas (see `references/gotchas.md`)

- **Agent run-as has View All** — Agent sees cross-owner records even when user shouldn't.
- **Conversation retention unset** — GDPR subject request cannot locate conversation logs.
- **Shield Event Monitoring not streamed** — Agent anomaly is invisible to SOC.

## Top LLM Anti-Patterns (full list in `references/llm-anti-patterns.md`)

- Running the agent as the invoking user by default — privileged reviewers leak permissions upward.
- Generic 'UpdateRecord' actions — any field becomes attack surface.
- Skipping DMO classification — grounding becomes a data-leak vector.

## Official Sources Used

- Agentforce Developer Guide — https://developer.salesforce.com/docs/einstein/genai/guide/agentforce.html
- Einstein Trust Layer — https://help.salesforce.com/s/articleView?id=sf.generative_ai_trust_layer.htm
- Invocable Actions (Apex) — https://developer.salesforce.com/docs/atlas.en-us.apexref.meta/apexref/apex_classes_invocable_action.htm
- Agentforce Testing Center — https://help.salesforce.com/s/articleView?id=sf.agentforce_testing_center.htm