Verifying Remediation Intent: Toward Trustworthy Autonomous Compliance Remediation for Critical Infrastructure

Automated tools can now detect compliance misconfigurations in infrastructure-as-code and propose, or apply, remediations. The unsolved problem is not detection or generation but verification: confirming that a proposed change satisfies the intent of the control it claims to fix, rather than merely passing a syntactic validator. This paper characterizes a specific failure mode, the silent false-fix: a change that is structurally valid, applies cleanly, and is reported as successful while the control remains violated. We argue that this failure is the central barrier to trusting autonomous remediation on operational technology (OT) and other critical infrastructure, where verification by replay is impossible and a wrong change has physical consequences. We introduce RIV-Bench, an open and reproducible benchmark that operationalizes the distinction between passing a linter and satisfying a control's intent, and report a naive lint-only baseline that approves every structurally valid but non-compliant remediation in a labeled fixture set spanning the CIS Benchmarks (AWS, Azure, GCP) and SOC 2, for a false-fix rate of 0.71. We then situate verification within a broader research program for trustworthy autonomous remediation, defined by three properties an AI-proposed change must satisfy and demonstrate: correctness against control intent, authenticated provenance, and bounded behavior. RIV-Bench and the analysis here are method-neutral: they define and measure the problem rather than prescribe a proprietary solution.