meta_pixel
Tapesearch Logo
Log in
The John Batchelor Show

S8 Ep157: DeepSeek's Security Backdoors — Jack Burnham — Burnham reports that the Chinese AI model DeepSeekgenerates code containing severe security vulnerabilities when queried regarding Chinese Communist Party-sensitive topics including Tibet, Xinjiang, and Taiwa

The John Batchelor Show

John Batchelor

News, Arts, Books, Society & Culture

4.52.8K Ratings

🗓️ 3 December 2025

⏱️ 13 minutes

🧾️ Download transcript

Summary

DeepSeek's Security BackdoorsJack BurnhamBurnham reports that the Chinese AI model DeepSeekgenerates code containing severe security vulnerabilities when queried regarding Chinese Communist Party-sensitive topics including Tibet, Xinjiang, and Taiwan, demonstrating that the model contains embedded political surveillance and control mechanisms. Burnham characterizes DeepSeek as possessing a "split personality": technical competence in general programming tasks combined with sophisticated political filtering and censorship capabilities. Burnhamrecommends urgent prohibition of such Chinese AI models from American critical infrastructure, government systems, and defense networks due to inherent security risks and embedded espionage capabilities.
1956

Transcript

Click on a timestamp to play from that location

0:00.0

This is CBS Eye on the World. Here's John Batchelor.

0:12.3

This is CBS, I on the World. I'm John Batchelor. Deep Seek, gaining attention around the world when it was introduced by the People's Republic of China's media,

0:23.7

Deepseek is said to be a program running cheaper, faster, just as good as the expensive large language models in America

0:31.0

and now in competition with each other, anthropic, open chat, Gemini.

0:37.4

I welcome Jack Burnham, a research fellow at the Foundation for

0:41.3

the Defense of Democracy, who guides me to a new revelation about Deep Sea. I've been watching the

0:47.5

hardware and claims that, for example, we don't use Nvidia chips or claims that it's cheaper than the runs through, the large language model run through is developing over at Claude for Anthropic.

1:01.9

I had not looked at the software and Jack now guides me to why that's important. Jack, a very good evening to you.

1:08.4

DeepSeek has many iterations, R1, R2, R3, for example.

1:14.6

However, what is important is the quality of information you get from asking questions for

1:20.3

deep seek. What have you discovered thanks to research by a very important crowd strike security firm watching the internet. Good evening to you.

1:31.4

Thank you for having me. I think what's interesting about what crowds right discovered is that

1:37.0

while deep seek when used normally as a tool in order to, for example, solve different coding challenges to produce apps and

1:46.4

websites, when asked to do so in a neutral way, often does so very well, on par with some of its

1:52.1

larger American competitors like Anthropic or ChatGPT from OpenAI. On the other hand,

1:58.8

what's really worrying about what CrowdStrike discovered is that

2:02.3

when CrowdStrike analysts introduced terms that Beijing deems politically sensitive,

2:09.0

such as references to Uyghurs for a persecuted minority or to Tibet, then code that's developed using those terms often contains significant security vulnerabilities.

2:23.0

Now, for example, CrowdStrike asked Deepseek to create an app that would allow Uyghurs to connect with one another online.

2:32.1

And the code that Deepseep produced in response to that

2:35.0

contained significant security vulnerabilities

2:37.0

that had backdoors that would allow authorities,

...

Please login to see the full transcript.

Disclaimer: The podcast and artwork embedded on this page are from John Batchelor, and are the property of its owner and not affiliated with or endorsed by Tapesearch.

Generated transcripts are the property of John Batchelor and are distributed freely under the Fair Use doctrine. Transcripts generated by Tapesearch are not guaranteed to be accurate.

Copyright © Tapesearch 2026.