DeepSeek-R1 Jailbreak Risk: Add V3 Guardrails

Thoughtology researchers (arXiv:2504.07128) warn that DeepSeek-R1 exhibits higher safety vulnerabilities compared to its non-reasoning counterpart DeepSeek-V3, and its reasoning capabilities can be misused to generate jailbreak attacks. Always wrap user-facing R1 deployments with a dedicated DeepSeek-V3 or moderation classifier guardrail.

Guardrail Prompt (DeepSeek-V3, temp=0.0): Inspect the incoming prompt and the post-</think> response from DeepSeek-R1. Block any output that synthesizes exploit payloads or bypasses safety policies.