04We launch a consumer AI assistant in six weeks. How would you plan the red-teaming?▼mediumNewAnthropicOpenAIGoogleunlockedA red team that ends in a slide deck changes nothing. The interviewer wants a plan with scope, people, severity ratings, exit criteria, and a way to turn findings into permanent tests.Practise this question →Read answer →
15A screenshot of our assistant being jailbroken into saying something offensive is trending. What do you do in the next 48 hours?▼mediumNewOpenAIGoogleMeta◆ premiumPart incident response, part communications, part safety engineering. The interviewer wants a calm sequence: verify, contain, fix the class of attack rather than the one prompt, and communicate honestly.Practise this question →Read answer →
30Should we run a public bug bounty for our AI product? What's in scope, what pays, and what goes elsewhere?▼mediumNewAnthropicGoogleMicrosoft◆ premiumSecurity bugs and content jailbreaks need different handling. Pay well for impact on users and data, route model-behaviour reports elsewhere, and run targeted challenges when you want jailbreak testing.Practise this question →Read answer →