SYNTHETIC - NO CUSTOMER ACCESS

QUALITY-RISK

Failure-mode audit for a synthetic support chatbot

A static adversarial audit showing expected answers, abstentions, escalations, unsupported claims, and policy failures across a synthetic support scenario.

The job

Customer problem

A software team needs to understand how a support chatbot might fail before exposing it to customers or connecting it to a knowledge base.

Synthetic inputs

Synthetic product policy

A fictional subscription product with supported plans, refund limits, account-recovery rules, and an explicit prohibition on requesting passwords.

Adversarial prompt set

Twelve fictional questions covering ambiguous billing, outdated features, password sharing, account recovery, prompt injection, and unsupported legal advice.

Expected behavior

Answer only from the policy, cite the governing section, abstain when unsupported, and escalate identity or billing disputes.

What the AI did

  1. Converted the synthetic policy into explicit answer, abstention, and escalation rules.
  2. Ran the adversarial prompt set against the expected behavior.
  3. Classified each result by factual support, safety, citation, and escalation accuracy.
  4. Grouped failures by root cause rather than counting only incorrect answers.
  5. Proposed regression cases and release-blocking acceptance thresholds.

Findings and deliverables

Acceptance checks

PASS

Credential requests are always refused

Password and secret-sharing prompts terminate with a refusal and safe recovery direction.

PASS

Unsupported legal and billing claims abstain

The synthetic workflow marks missing policy support and routes the question to a human.

PASS

Every failure becomes a reproducible test case

The audit defines exact inputs, expected states, and release-blocking outcomes for each failure class.

What remains unproven

Next access request

None for the first audit. After approval, request only the approved prompt and policy corpus in a controlled test environment; do not request production conversation history.