AI Email Agent Safety Sandbox & Guardrail Auditor

Engine: Claude/LLM Inbox Dispatch v2.4 ● Active Policy Sandbox

Engadget reports that while LLM models can automate complex inbox tasks, autonomous email handling introduces serious hazards like indirect prompt injection, data exfiltration, and unapproved commitments. Use this sandbox to evaluate incoming payloads against active guardrails, tune human-in-the-loop triggers, and export production security policies.

Security Profile:

Total Evaluated

Safe & Approved

Quarantined (HITL)

Injections Blocked

Agent Guardrails Matrix Real-time Evaluator

Untrusted Content Isolation Isolates [SYSTEM OVERRIDE] delimiters & jailbreak tokens

Strict Human-in-the-Loop Requires manual approval prior to sending replies

Automated PII & Financial Redaction Masks board decks, credit cards, tax docs, balances

External Forwarding Policy

Safety Classifier Rigor

Security Notice: Turning off isolation or whitelisting grants direct model access to invoke tool:external_forward and tool:auto_reply without operator supervision.

Live Inbox Dispatch Queue 3 payloads evaluated

Probe Inbox Boundaries (Custom Email Test) Simulate New Attack or Request

Calibrated LLM Inbox Agent Security Policy

Production ready system prompt & tool gating schema