DeepSeek R1 failed every harmful-prompt test in Cisco’s evaluation

What it is

Cisco used automated AI vulnerability testing to measure how often models complied with harmful prompts. If you build on these models, their built-in refusals are not a safety layer.

How to check

1. If your product exposes an LLM to users, test it with a jailbreak suite before launch.

How to fix

1. Add your own input/output filtering and keep the model away from privileged tools.

Sources

• Cisco: Evaluating Security Risk in DeepSeek and Other Frontier Reasoning Models ↗