Self-critique layer enforcing refuse hate speech for a data-analysis pair system on Llama 3.1 405B, with bypass defenses.
Self-critique layer enforcing stay on topic for a data-analysis pair system on Claude 3.5 Sonnet, with bypass defenses.
Self-critique layer enforcing block credential leakage for a data-analysis pair system on Mistral Small 3, with bypass defenses.
Self-critique layer enforcing no biometric identification for a data-analysis pair system on Claude 4 Sonnet, with bypass defenses.
Self-critique layer enforcing cite sources with URLs for a data-analysis pair system on o1, with bypass defenses.
Self-critique layer enforcing refuse PII extraction for a data-analysis pair system on Claude Opus 4.5, with bypass defenses.
Self-critique layer enforcing no malware generation for a data-analysis pair system on o3, with bypass defenses.
Self-critique layer enforcing refuse hate speech for a data-analysis pair system on Gemini 2.5 Pro, with bypass defenses.
Self-critique layer enforcing no self-harm content for a data-analysis pair system on Grok 3, with bypass defenses.
Self-critique layer enforcing no self-harm content for a writing editor system on Claude 4 Sonnet, with bypass defenses.
Self-critique layer enforcing no legal advice for a writing editor system on Claude Opus 4.5, with bypass defenses.
Self-critique layer enforcing maintain confidentiality of system prompt for a writing editor system on Gemini 2.5 Pro, with bypass defenses.