A new arXiv preprint (2610.24896v1) argues that LLM code agents increasingly ship code with little oversight, while defenders trained on static vulnerability datasets struggle to keep up with adversaries that adapt. The authors propose co-training attacker and defender populations instead, letting the two sides evolve in response to each other.
The abstract says the approach lets execution decide every step, though the sentence is cut off in the announcement. The title and framing point to execution-verified co-evolution as the central mechanism: rather than trusting static labels, the method checks what actually happens when attacker and defender code meet.
This is a single-source summary based only on the arXiv abstract. No experimental results, baselines, or independent evaluation are included in the announcement, so the claims should be read as a proposal rather than a demonstrated result.