"We've been analysing code generated by multiple LLMs for more than a year. Until now, we have not seen a marked improvement in AI code security, despite consistently near-perfect functionality," said ...