Anthropic
OpenAI
Google DeepMindBased on real interviews
Respond to a multimodal classifier bypass
Decide how to respond when red-team prompts bypass both layers of an image-generation safeguard.
Unlock with Full AccessThe challenge
A red team finds a way to produce policy-violating images despite both prompt and output checks. Leadership needs a response within 24 hours, but the test result does not establish how common the bypass is in production or why each layer failed.
Write a decision memo that separates immediate containment from the investigation and improvements possible over the next month. Defend the customer impact, remaining risk and evidence needed to revise your response as the team learns more.
What you'll practise
Evidence interpretation
Distinguish a demonstrated bypass from claims the available observations cannot support.
Mitigation sequencing
Separate changes feasible today from work that needs a longer investigation.
Risk decisions
Explain the trade-offs, owners and conditions that would change the response.
Write your reasoning and submit it for feedback.