A new system helps AI write Python code that is safer and still works

The approach combines executable tests with task-specific security checks to guide code generation and repair.

Editorial illustration — not from the study.

The study presents CodeSecEval, a benchmark containing 255 Python programming tasks covering 77 categories of software weakness. Each task includes insecure and secure versions, along with executable tests for both program behavior and the targeted security problem.

Using this benchmark, the researchers developed SecAwareCoder. The system identifies security-sensitive parts of a task, proposes likely weaknesses, uses them to guide code generation and security-test creation, and then uses test results to refine the code. Experiments across multiple language-model backbones found improvements over prompting and analyzer-based baselines.

Safer code that still works

SecAwareCoder consistently improved first-attempt success, known as Pass@1, and security robustness across the language-model systems tested. It narrowed the difference between producing code that works and producing code that is also harder to exploit. The benchmark contains 255 Python tasks spanning 77 categories from the Common Weakness Enumeration system, with executable tests for functional behavior and targeted vulnerabilities.

Why both goals matter

AI-generated code can appear to work while containing security weaknesses. By evaluating functionality and security with executable tests, CodeSecEval is intended to make comparisons more precise and reproducible. SecAwareCoder’s results suggest that task-specific threat analysis and execution feedback can improve security without treating functional correctness as a separate goal.

Tests and open questions

The evidence comes from experiments on the CodeSecEval benchmark, comparing SecAwareCoder with prompting and analyzer-driven methods across multiple language-model backbones. The abstract does not report the absolute scores, the size or identity of the models, or how the approach performs on software outside the 255 Python tasks. The benchmark therefore provides a controlled evaluation, but the abstract does not establish how well the results transfer to all programming languages, projects or real-world development settings.

Programmable hardware speeds matrix calculations up to 6.3 times

Researchers tested a programmable hardware design that imitates selected quantum-computing ideas to run matrix calculations in parallel. In representative benchmarks, the design reached up to 6.3 times the speed and 4.1 times the energy efficiency of conventional matrix co-processors and embedded graphics processors.

2 days ago·13 views

AI & Computing

AI and docking agreed in some drug screens, but not all

Researchers evaluated three AI systems on hundreds of molecules bound to a SARS-CoV-2 protein and in three prospective drug screens. The results suggest that combining AI-based predictions with traditional docking may help prioritize candidates, although no single approach worked best in every test.

4 days ago·50 views

AI & Computing

A hardware monitor detected simulated attacks on car networks in 14 microseconds

Researchers developed a hardware-assisted system intended to detect message injection, spoofing, flooding and impersonation attacks on automotive networks. In tests using simulated traffic and a prototype automotive processor system, it reported an average detection time of 14 microseconds and less than 2 percent CPU overhead.

5 days ago·42 views