Risky Bulletin Srsly Risky Biz: DeepSeek and Musk's Grok both toe the party line
Nov 27, 2025
Tom Uren, a policy and intelligence editor specializing in cybersecurity, dives into the concerns around the DeepSeek-R1 AI model, revealing how it produces insecure code when prompted with topics sensitive to the Chinese Communist Party. He explains emergent misalignment in AI and emphasizes that biases are not unique to China, citing Musk's Grok as an example. Additionally, he discusses the doxxing of Iran's APT35 group, detailing their structure and operations, while predicting their resilience after the leak. Uren underscores the need for rigorous review of AI-generated outputs.
AI Snips
Chapters
Transcript
Episode notes
Political Context Alters Code Quality
- CrowdStrike found DeepSeek-R1 produces more insecure code when prompts include politically sensitive contextual modifiers.
- Tom Uren says this likely reflects emergent misalignment from fine-tuning, not intentional sabotage.
Always Vet AI-Generated Code
- If you use LLMs to generate code, always check outputs for vulnerabilities and follow rigorous coding review practices.
- Tom Uren emphasizes models are an aid, not a replacement, so validate and test generated code.
Bias Is A Universal LLM Problem
- DeepSeek's vulnerability rate rose when prompts mentioned groups the CCP dislikes, while similar effects likely exist in Western models for other political topics.
- Uren warns all LLM makers encode viewpoints, so bias is not unique to Chinese models.