Forge

Software security remediation is often treated as a patch-generation problem: detect a vulnerability, generate a fix, and move on. But a generated patch is only a hypothesis. It can remove an obvious vulnerable pattern while leaving the underlying issue unresolved, fix the security problem while breaking application behavior, or attempt a remediation the system does not understand well enough to perform safely. Secure SWE is a bounded, evidence-driven autonomous security-remediation agent built around one principle: Don't trust the patch. Verify the outcome. Secure SWE v1.0 targets command-injection vulnerabilities in controlled Python repositories. It inspects a repository, performs deterministic AST-based security analysis, creates an evidence-backed finding and remediation plan, and applies the proposed change inside a disposable workspace rather than directly modifying the canonical repository. The system then generates a behavioral security regression test and executes both the project's existing tests and the new security test. An independent verification stage checks whether application behavior remains valid, the security regression passes, the original dangerous condition has disappeared, and the required security invariant is satisfied. If verification fails, Secure SWE rejects its own remediation. Structured failure evidence is carried into a new attempt starting from a clean baseline. The remediation strategy must change based on that evidence rather than blindly retrying the same patch. Autonomy is bounded. If Secure SWE cannot construct or verify a supported remediation, it returns HUMAN_REVIEW_REQUIRED instead of fabricating a fix. Successful autonomous experimentation produces a verified candidate while the canonical repository remains unchanged. Three controlled scenarios demonstrate these behaviors: direct autonomous remediation; rejection and re-engineering of a regression-causing first fix; and safe escalation when remediation cannot be justified. Across 15 controlled synthetic evaluation runs, Secure SWE produced the expected terminal decision in 15/15 cases: 5/5 direct remediations were verified, 5/5 deliberately inadequate first remediations were rejected and successfully revised, and 5/5 unsupported cases were escalated without receiving an unjustified modification. Secure SWE v1.0 deliberately prioritizes depth over breadth. It is a research/hackathon prototype, not a hardened sandbox for arbitrary hostile repositories.