Interview, Fireside Chat
How to use DeepSeek safely
- Strategic Recommendation: Enterprise and infrastructure deployments of DeepSeek are advised against until the technology stabilizes and known security ambiguities are resolved; the model is characterized as volatile for production use.
- Security Vulnerability: DeepSeek exhibits significantly weaker jailbreak resistance compared to GPT, performing approximately 20% worse on benchmark tests and matching the security maturity of GPT-3.5 regarding susceptibility to basic injections and zero-day attacks.
- Infrastructure Risk: The underlying infrastructure used to train and operate DeepSeek is considered insecure, necessitating extensive hardening for any system integrating the model to prevent pivoting attacks.
- Deployment Advice: If deployment is deemed necessary immediately, it should be restricted to non-end-user-facing use cases to mitigate the operational burden of securing basic vulnerabilities.
- Censorship Scope: The model enforces heavy censorship on Chinese political topics (e.g., Tiananmen Square, Taiwan), with 85% of such topics in test sets triggering hard refusals or mandatory alignment with CCP narratives.
- Version Consistency: Censorship protocols are identical across the China-hosted version, US-hosted versions, and the locally runnable open-source variant, as the base model itself contains the guardrails rather than relying solely on client-side filtering.
- Unknown Risks: Beyond overt censorship, significant uncertainty remains regarding potential subtle manipulations, backdoors, or hidden triggers inserted by the Chinese government to bypass guardrails.
- Comparative Analysis: Western models (e.g., Claude, GPT) exhibit similar censorship behaviors regarding sensitive Chinese political topics (censoring ~40% of such topics) compared to DeepSeek's 85%, challenging the narrative that only Chinese models suppress specific geopolitical content.
- Model Performance: DeepSeek is noted for operational inefficiencies, including slow inference speeds, verbose output, and errors such as the random insertion of Chinese characters in responses.
- Future Alternatives: Experts anticipate an imminent release of a stable, open-source model utilizing the same reasoning reinforcement learning techniques without the associated security or political ambiguity.
- Hosted Model Distinction: The only functional difference between the China-hosted and open-source versions is an additional post-generation output filter on the China-hosted version that clears sensitive content on the client side.