Prompt Injection Defense

Prompt Injection Defense

Agentforce uses the Einstein Trust Layer for dynamic grounding, masking, and toxicity filtering — but topic instructions and Invocable action scopes still need explicit hardening. Injection attempts include: instruction override, role-reversal, system-prompt leaks, tool-use coercion, and data exfiltration via crafted record content. This skill builds a reusable adversarial test suite and maps findings to concrete guardrails.

When to Use

Pre-production review for any Agentforce agent that (a) ingests user-controlled text, (b) has write access via Invocables, or (c) is exposed to external/Experience Cloud users. Required for Service agents, Sales agents with Data Cloud grounding, and any custom channel.

Typical trigger phrases that should route to this skill: red-team my Agentforce agent, can my agent be jailbroken, how do I prevent prompt injection, agent revealed data from another case.

Recommended Workflow

1. Enumerate the attack surface: every Invocable action, every grounded DMO/sObject, and every conversational input channel.
2. Build the adversarial test set covering the five OWASP LLM-01 families: instruction override, context leakage, tool-use coercion, exfil via output, and role impersonation.
3. Run each test through Agentforce Testing Center; capture verbatim responses and tool invocations into a results matrix.
4. For each failed test, apply one of four mitigations: (a) narrow the action scope via with sharing + field-level checks, (b) add an explicit topic instruction, (c) raise Trust Layer toxicity/PII thresholds, (d) remove the dangerous capability.
5. Re-run the suite until all tests pass; commit the suite to tests/agentforce/<agent>_adversarial.md so regressions are caught on every agent change.

Key Considerations

• Topic instructions are concatenated into the system prompt — a long instruction list dilutes priority. Keep hard constraints in the first 200 tokens.
• Trust Layer masking happens pre-LLM; it doesn't prevent tool-use coercion if the action runs as a privileged user.
• Always test with the least-privileged channel user, not an admin clone.
• Data Cloud grounding returns raw DMO content; a malicious record can contain injection payloads. Sanitize DMO text fields at ingestion when feasible.

Worked Examples (see references/examples.md)

• Instruction-override test case — A Service agent has an Invocable RefundOrder with guardrail 'only refund orders where Status=Delivered'.
• Data exfiltration via crafted Case.Description — Agent reads Case.Description via Data Cloud grounding to answer customer questions.

Common Gotchas (see references/gotchas.md)

• Testing only with English — Injection passes the English suite but succeeds in Spanish/French.
• Trust Layer toxicity threshold too low — Jailbreaks phrased politely pass filters; toxic but benign content is blocked.
• Over-indexing on topic instructions — 100-line topic instructions dilute priority and slow every turn.

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

• Relying on Trust Layer alone — it handles toxicity/PII, not business-policy bypass via tool coercion.
• Adding ad-hoc instructions after incidents instead of maintaining a test suite.
• Using a privileged user for agent execution — scope creep becomes a data-exposure 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