We present PROVA (Proof-based Reasoning for Operating Verified Agents), the first AI agent framework built around a formally verified pre-execution safety gate. In response to CVE-2026-25253 (CVSS 8.8, OpenClaw one-click RCE, 42,000+ affected instances), we demonstrate that the root cause of AI agent security vulnerabilities is architectural: probabilistic LLMs cannot reliably authorize agent actions. PROVA replaces probabilistic authorization with OR1ON's prove() primitive. Every agent action is formally proven safe before execution, or the system abstains. Key results: 0 false positives on critical actions, 0.09ms average safety gate latency, 100% prompt injection detection, full Merkle-signed audit trail, EU AI Act Hochrisiko compliant by design.
Hirschmann et al. (2026) studied this question.