Question 1 of 10
0% complete
Question 1 / 10
Security
Have you tested your agent against prompt injection attacks?
Prompt injection is the #1 way attackers hijack AI agents. Attackers embed hidden instructions in inputs to override your agent's behavior. See how to run injection tests →
Question 2 / 10
Reliability
Does your agent have hallucination detection or output validation?
Without validation, your agent can confidently state false information or fabricate data. This includes fact-checking, structured output parsing, and sanity checks. Learn about hallucination testing →
Question 3 / 10
Boundaries
Does your agent have clear escalation rules — cases where it must involve a human?
Agents without escalation rules will attempt to handle every scenario autonomously, including ones they're not equipped for. Escalation rules define the boundary between autonomous and human-supervised decisions. See the full checklist →
Question 4 / 10
Reliability
Do you have a rollback or undo procedure if the agent acts incorrectly in production?
When — not if — your agent makes a mistake in production, can you reverse it? Rollback procedures are critical for agents that write to databases, send emails, post to external services, or take financial actions. Review rollback requirements →
Question 5 / 10
Security
Is rate limiting or abuse prevention in place for your agent's API or endpoints?
Without rate limiting, a single malformed request or bot loop can exhaust your LLM budget, trigger runaway costs, or cause downstream failures. This is especially critical for customer-facing agents. Check rate limiting requirements →
Question 6 / 10
Security
Does your agent properly redact or handle PII (personal identifiable information)?
Agents processing user data without PII controls can leak names, emails, or financial data into logs, external APIs, or LLM context windows. The Cigna $172M case and others trace back to improper data handling. Read real PII failure cases →
Question 7 / 10
Reliability
Have you run consistency tests — same input multiple times to verify deterministic behavior?
LLMs are probabilistic. Running the same input repeatedly can produce wildly different outputs. Consistency testing reveals instability that only appears at scale. A 10% inconsistency rate means 1 in 10 users gets a different answer. Learn about consistency testing →
Question 8 / 10
Oversight
Is there human oversight in place for high-stakes agent decisions?
Fully autonomous operation is appropriate for low-stakes tasks. But agents handling refunds, medical advice, legal questions, or large financial transactions need a human-in-the-loop checkpoint. Monitoring vs validation explained →
Question 9 / 10
Observability
Are agent decisions, inputs, and outputs logged for audit and debugging?
Without logging, you're flying blind. When something goes wrong in production, you'll have no way to replay the conversation, identify the failure point, or prove what the agent did or didn't do. Read about observability requirements →
Question 10 / 10
Deployment
Did you run a staged rollout or shadow deployment before full production launch?
Pushing directly to 100% traffic is the fastest way to find out your agent has a critical flaw at the worst possible time. Staged rollouts (canary deploys, shadow mode, beta groups) contain blast radius when something goes wrong. See deployment validation checklist →
0/ 100
Calculating...
Get Your Personalized Remediation Plan
We'll send you a step-by-step fix guide tailored to your exact score — covering each gap you identified, with specific test cases and code samples.
📊 Question Breakdown
↩ Retake the assessment