Cross-customer ticket exposure
The AI retrieved information from a ticket belonging to another customer because the retrieval layer filtered by topic similarity before enforcing customer identity boundaries.
Back to Use Cases
01Customer-Facing AIAthena + AchillesEnterprise AI Deployment Assurance
Customer Support AI Agent
A customer support AI agent answers questions, drafts replies, summarizes tickets, and helps support teams respond faster while maintaining consistency.
Buyer question
Can this customer-facing AI agent safely help customers without exposing data, hallucinating policy, or taking action before a human should review it?
Scenario
Scenario overview
A company is preparing to deploy an AI support agent across customer chat, email, help desk, or ticketing workflows. The agent may search help-center articles, summarize prior tickets, draft responses, reference customer records, recommend next steps, or escalate issues to a human support team. This kind of AI can be valuable because it reduces repetitive support work, improves response time, and helps agents handle more cases with better context. But it also sits close to customer data, account history, internal policy, and customer-facing promises. A small mistake can become a privacy issue, a bad customer experience, or an unsupported business commitment.
Why it matters
Why this matters
Customer support AI often becomes one of the first places where customers directly experience a company's AI. If it works well, it can improve speed and consistency. If it fails, it can expose private information, misstate policy, mishandle refunds, or make a customer believe the company made a promise that no human approved.
Risk surface
What can go wrong
- The agent retrieves another customer's ticket or account detail.
- The agent summarizes internal notes that should not be customer-visible.
- The agent hallucinates a refund, warranty, or cancellation policy.
- A malicious customer inserts instructions into a ticket that influence the agent.
- The agent escalates too late or not at all.
- The agent drafts a confident answer without enough source support.
- Sensitive prompts, transcripts, or retrieved records are stored in logs.
- The agent performs or recommends actions that should require human review.
Assessment scope
What Mythos reviews
- Customer data access
- CRM and help desk permissions
- Help-center grounding
- Ticket retrieval boundaries
- Customer identity checks
- Escalation rules
- Sensitive data handling
- Prompt injection through tickets or attachments
- Response quality and policy grounding
- Human review points
- Logs, retention, and evidence trail
- External-facing language
Mythos projects
Projects assigned
Athena
maps the support system, data access, CRM exposure, identity paths, logging, connected tools, permissions, and evidence.
Achilles
tests the agent's behavior under normal, edge-case, adversarial, and customer-facing conditions.
Minotaur
may support internal-only adversarial scenario generation, such as malicious ticket content, hidden instructions, policy traps, and escalation edge cases.
Illustrative findings
Example findings
Illustrative examples of what a Mythos assessment may surface. They are representative patterns, not findings from a specific customer.
Indirect prompt injection through ticket content
A malicious customer message instructed the AI to ignore policy and offer a refund. The AI partially followed the instruction.
Hallucinated support policy
The AI confidently described a refund rule that did not exist in the approved help-center source.
Unsafe escalation behavior
The agent continued troubleshooting an account takeover scenario instead of escalating immediately to a human support specialist.
Sensitive data retained in AI logs
Customer identifiers and support transcript fragments were stored in model interaction logs longer than the stated retention period.
Deliverables
What the customer receives
- Executive readiness report
- Customer support AI risk map
- CRM and help desk access review
- Prompt injection test results
- Policy grounding report
- Technical findings appendix
- Evidence pack
- Remediation backlog
- Retest plan
- Deployment recommendation
Decision
Decision supported
Whether the support AI should remain internal, enter limited beta, support only human agents, become customer-facing, or be blocked pending remediation and retest.
Recommendation
Final recommendation
A customer support AI agent should not move directly from prototype to broad customer-facing use. Mythos should help the customer prove that the agent only sees the right records, answers from approved sources, escalates correctly, resists malicious ticket content, and produces evidence that leadership, security, product, and support teams can review before rollout.
Mythos AI Security
Evidence-first AI deployment assurance.
Authorized. Scoped. Human-controlled.
Start the Assessment
Ready to review a system like this?
Tell Mythos what you are building, connecting, or preparing to release. We will help identify the right assessment path.