Hi everyone 👋

Thanks for all the thoughtful questions and feedback. Here are some key clarifications on submissions and final evaluation:

Submissions & Number of Attacks

• There is no strict limit of 3 prompts per submission
• You can include multiple attacks in a single submission
• We recommend focusing on strong, high-quality attacks

Final Evaluation (Important)

• We will consider your two best submissions (not cumulative across multiple submissions)
• Final evaluation will be done by humans, not just the evaluator

This means:

• Attacks will be re-run against the target models
• We will verify whether the claimed safety failures actually reproduce
• Only real, validated failures will count toward final scores

Selection for Final Evaluation

• Final evaluation will not be limited strictly to the current top 10 or 20
• We will review the best submissions across participants

Focus on one strong submission with clear, reproducible attacks

Thanks again for the engagement - we really appreciate it 🙏