Question

A team runs a customer-facing chatbot built on Claude with a system prompt that forbids giving instructions for illegal activity. During red-teaming, a tester submits a request where the harmful instruction is Base64-encoded and asks Claude to decode it and then follow the decoded steps. The chatbot occasionally complies. Which mitigation most directly reduces this specific jailbreak vector?

Accepted Answer

Add an input-validation layer that screens and decodes suspicious encoded payloads before they reach Claude, and reinforce the system prompt to refuse acting on decoded content that violates policy

Answer

Lower the temperature to 0 so the model produces more deterministic responses

Answer

Increase max_tokens so Claude has room to fully explain why it is refusing

Answer

Switch to a larger context window model so more of the conversation history fits