Agents
MACGen splits secure code generation into planner, security advisor, coder and reviewer passing structured artifacts instead of shared dialogue
arXiv 2608.25457 (26 Aug) treats secure code generation as the multi-objective problem it is, functional correctness plus security, and argues the two common fixes both fail: retrieved security guidelines leave the generator to translate generic advice into a task-specific implementation, and shared-dialogue multi-agent feedback blurs role boundaries and bloats context. MACGen has a security advisor identify likely CWEs and synthesize task-specific guidelines, then gives each downstream agent only structured artifacts from upstream rather than full history. It improves F&S@1 over direct prompting by 19.61 percentage points on CWEval and 10.57 on BaxBench.
Source
↳ Follow the thread