Ian Webster (Promptfoo) on DeepSeek’s Security Vulnerabilities Ian Webster, founder of Promptfoo, joins a16z partner Joel de la Garza to break down the security risks embedded within DeepSeek’s reasoning model. As generative AI systems become more powerful, they also become more susceptible to attack. Ian explains how vulnerabilities like jailbreaks, backdoors, and model censorship can be exploited—and what developers and enterprises can do to defend against them. He also shares insights into how AI security testing is evolving, why transparency in model training matters, and what lessons companies can take from past security breaches to safeguard the next wave of AI applications. Learn more: What Are the Security Risks of Deploying DeepSeek-R1? - https://www.promptfoo.dev/blog/deepseek-redteam/ Follow everybody on social media: Ian Webster - https://x.com/iwebst Joel de la Garza - https://www.linkedin.com/in/3448827723723234/ Check out everything a16z is doing with artificial intelligence, including articles, projects, and more podcasts, here: https://a16z.com/ai/ 01:11 - DeepSeek: The Golden Age of AI or an existential threat? 02:18 - Red team testing, prompt injections, jail brakes - adversarial techniques 02:48 - Speech limitations 04:14 - Maturity and complexity of DeepSeek vs. other models 05:36 - Anything you build on top of DeepSeek will be subject to its insecurities 06:12 - Hosted model from China vs. open source/running locally 07:46 - DeepSeek benchmark on politically sensitive topics 08:54 - Western censorship vs. DeepSeek censorship 12:38 - How can we use it safely? Protecting infrastructure 14:09 - Wait for a more trusted source to run locally?