What evidence should gate an AI-generated patch?

Which independent checks change the decision to accept an AI-generated patch, and what do they cost?

assistant-draftedVulnerability repairEvaluation validity

Signals

Contested

DARPA's counted competition patches were qualified by PatchBench, where agents from top AIxCC teams lose much of their solve rate under stronger validation; the competition review also found many fully validated baseline-agent patches semantically wrong.

Incidents outpace defenses

Discovery outpaces patching, which raises pressure to accept patches quickly.

Fide agenda

Directly serves FID-088.

Why it matters

Maintainers are being asked to accept machine-written fixes faster than they can review them. A plausible patch that passes the original test can still leave the flaw open.

Hypothesis

A small set of independent security and regression checks catches most false acceptances at modest cost.

A first study

Take released AIxCC or PatchBench repairs, apply the original check and independent checks separately, and compare acceptance decisions and review cost.

Controls it would need

Checks defined before evaluation; held-out repairs; domain review to separate a failed repair from an invalid check.

What it could and could not claim

Finite checks give bounded evidence, not proof that software is secure.