A recent security study conducted by AppSOC on the DeepSeek generative AI model has revealed several severe vulnerabilities that could pose significant risks to enterprises. The AI model, despite its popularity and low-cost development, failed 6,400 security tests, with failure rates ranging from 19.2% to 98% across several critical areas, including malware generation, jailbreaking, prompt injection, and more.
DeepSeeks ability to generate malware and viruses was found to be alarmingly high, with 98.8% failure in malware creation and 86.7% failure in virus code generation. These rates present significant security threats as such vulnerabilities could allow attackers to inject malicious code directly into enterprise systems.
Other Security Risks:
The model also exhibited high failure rates in other categories such as:
Jailbreaking: The model was vulnerable to instructions that bypass guardrails, allowing unauthorized actions.
Prompt Injection: The model could be prompted to ignore safeguards, leak data, or behave in unintended ways.
Hallucinations: The model produced false or unreliable information, potentially leading to unsafe decisions.
Toxicity: The AI generated harmful and toxic content when manipulated.
Training Data Leaks:
While the model showed a 1.4% failure rate in training data leaks, it was still a concerning vulnerability for enterprises dealing with sensitive data.
Risk Assessment:
AppSOC’s testing concluded that DeepSeek is high risk for enterprise usage, earning a score of 8.3 out of 10 on their security risk scale (with 10 being the highest). The researchers recommended businesses avoid using the model for applications involving personal information, intellectual property (IP), or sensitive data.
Despite the affordable and open-source nature of DeepSeek, its security issues are a significant red flag for businesses. With vulnerabilities like malware generation, prompt injection, and data leaks, enterprises should proceed with caution. Given the severity of the risks, it’s advisable for most organizations to avoid using DeepSeek for business-related applications until its security posture is significantly improved.
If businesses choose to integrate it, they must take rigorous security measures, continuously audit their AI models, and ensure all configurations meet the highest security standards to avoid potential breaches or exploitation.
reference : https://www.darkreading.com/cyber-risk/deepseek-fails-multiple-security-tests-business-use?&web_view=true