Security audit
Security checks for vulnerabilities and agentic risk
The skill is a coherent ecommerce chatbot design guide, but it recommends full transcript sharing and automated customer-account actions without enough privacy and confirmation guardrails.
Review this before installing if you plan to turn the templates into a live chatbot. Add explicit customer confirmation for cancellations, refunds, replacements, return labels, and claims; avoid broad single-word intent triggers for order-data flows; and replace full-transcript sharing with a minimized, redacted handoff summary unless your privacy policy, consent flow, retention rules, and support access controls clearly allow more.
The markdown explicitly recommends escalating when a customer types "help me," which is a broad phrase commonly used in many benign contexts. Because this file is markdown and SQP-1 applies, this trigger is too unspecific and could cause unintended activation or escalation without clarifying scope or exclusions.
The skill recommends pre-filling escalation tickets with the full conversation transcript but provides no guidance on data minimization, redaction, retention, or consent. In an ecommerce support context, transcripts may contain order details, addresses, emails, payment-adjacent information, or other personal data, increasing privacy exposure and the risk of unauthorized access or over-collection.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
|---|---|---|
| Unhandled intent | Weekly review | Add to flow build queue |
| Low CSAT bot rating | <3/5 rating given | Review transcript; identify gap |
| Long session with no resolution | Session > 10 min, no confirmation | Flag for quality review |
## Escalation Handoff Protocol
The Order Status intent includes broad triggers like "track," "tracking," and "delivery," which can match many unrelated ecommerce questions and cause the bot to enter an account-specific flow prematurely. In this context, that can lead to unnecessary collection of order numbers or email addresses and misrouting of support requests, increasing privacy and workflow errors.
The Shipping Delay intent relies on ambiguous one-word triggers such as "late," "stuck," and especially "where," which are likely to fire on unrelated messages. Because the response template then instructs the system to look up order status, this increases the risk of improper intent activation, unnecessary access to order data, and confusing or privacy-impacting responses.
The warm handoff script promises that the entire conversation will be shared with another team without asking for consent or warning the user about what information will be transferred. In a customer service setting, conversations may contain personal, order, or payment-adjacent details, so automatic sharing can create unnecessary internal data exposure and privacy compliance issues.
The checklist explicitly requires escalation to pre-fill tickets with the full conversation transcript, but it includes no corresponding requirement to notify users, minimize sensitive data, or obtain appropriate consent where needed. In a customer-service chatbot context, transcripts often contain personal, financial, or order-related information, so silent forwarding to humans can create privacy and compliance risks if users do not expect broad internal sharing.
Skill's behavior or capabilities extend beyond its stated purpose. Scope creep allows an agent to perform actions unrelated to its documented functionality, increasing the attack surface.
9. **CSAT survey fatigue** — Sending a satisfaction survey after every bot interaction annoys customers. Survey 20–30% of interactions randomly, or only after escalation.
10. **Trying to handle everything at launch** — 5 well-designed flows that cover 70% of volume outperform 20 incomplete flows covering 100% of scenarios with gaps and dead ends.
## Resources
Automatically generating or emailing a prepaid label may involve using or transmitting customer contact and order data. The checklist describes this behavior but provides no warning or disclosure about the data use or email action.
No suspicious patterns detected.