A joint study by Stanford and MIT published in March 2026 analyzed over 2 million AI-generated code snippets and found that 14.3 percent contained at least one security vulnerability, compared to 9.
1 percent in human-written code for equivalent tasks.
