DeepMind Unveils CodeMender: An AI Agent for Software Code Security
DeepMind, a subsidiary of Google, has announced CodeMender—an artificial intelligence capable of detecting and resolving vulnerabilities in software code. CodeMender operates both reactively, by fixing detected errors, and proactively, by rewriting code to eliminate potential threats.
Developed based on the Gemini Deep Think models, the new AI agent combines methods of static and dynamic analysis, differential testing, and automatic verification of changes. In six months of testing, CodeMender has already made 72 fixes to open-source projects, all of which are reviewed by specialists before publication. This helps reduce the risk of regressions or new errors.
The technology used in CodeMender could significantly change the approach to ensuring software security. Traditionally, developers have used methods such as fuzzing for security checks; however, the new agent provides faster and more accurate vulnerability detection.
DeepMind plans to gradually expand access for developers to this tool and intends to publish the technical results of its research soon. This could have a major impact on the IT security industry by simplifying the process of updating and monitoring complex software systems.
Meanwhile, Google Research, in collaboration with DeepMind, has introduced VaultGemma—a language model with a billion parameters, trained using differential privacy technology. This signifies Google’s strategy to strengthen its position in the field of artificial intelligence, especially in light of the transition to so-called “superintelligence.”
| Feature | Benefits | Technologies |
|---|---|---|
| CodeMender | Automatic correction and proactive threat elimination | Gemini Deep Think, static and dynamic analysis |
| VaultGemma | Differential privacy, high accuracy | Billion parameters, DP technology |
