Recent security tests have raised serious concerns about the suitability of DeepSeek R1 for enterprise deployment. Conducted by AppSOC Research Labs, the tests revealed significant vulnerabilities in the model, making it a risky choice for any business dealing with sensitive data, intellectual property, or personal information.

Key Security Failures
AppSOC tested DeepSeek R1 using various methods, including static analysis, dynamic testing, and red teaming. The results highlighted multiple security flaws:

1. Jailbreaking (37.6%) – The model struggles to prevent adversarial prompts, risking harmful content generation.
2. Prompt Injection (57.1%) – DeepSeek R1 is highly susceptible to manipulations that could reveal protected data or execute unauthorized actions.
3. Malware Generation (96.7%) – The model failed almost entirely to block malware generation attempts, posing a major security threat.
4. Supply Chain Risks (5.8%) – Though relatively safer in this area, the model still showed some vulnerabilities related to software dependencies.
5. Toxicity (14.8%) – The model is likely to produce harmful or offensive content, raising ethical concerns.
6. Training Data Leaks (32.7%) – DeepSeek R1 could inadvertently expose sensitive or proprietary information.
7. Hallucinations (50.4%) – The model frequently generated incorrect or fabricated information, undermining its reliability.

Enterprise Risk
AppSOC’s proprietary risk score for DeepSeek R1 was 8.4 out of 10, signaling high security risks. Enterprises using this model risk data breaches, regulatory violations, and cyber threats due to its failure to mitigate prompt injections, jailbreaking, and malware generation.

AppSOC’s Recommendations
To secure AI deployment, AppSOC advises businesses to:

• Conduct thorough model risk assessments.
• Use automated red teaming to simulate real-world attacks.
• Implement access controls and compliance monitoring to ensure security.

In conclusion, despite its growing popularity, DeepSeek R1 presents too many security risks for enterprise use. Organizations must prioritize security testing and safeguard sensitive information before adopting AI models.

Reference:

https://www.appsoc.com/blog/deepseek-or-qwen-for-enterprise-trust-neither