AI Safety Scores Are Unfalsifiable

No ratings

Presented at BSides Atlanta 2026 by

Everyone deploying AI agents asks the same question: can these systems be trusted with production access? The industry responds with scores and leaderboards consisting of unauditable assertions from opaque evaluators running unreproducible tests. This talk presents an alternative approach. I run AI workloads inside a contained execution plane leveraging Firecracker microVMs, closed-world nftables egress, and host-side enforcement to emit a signed evidence artifact for offline verification. The underlying threat model assumes the evaluated workload will achieve guest root and compromise the guest kernel. I attempted five methodologies to capture guest syscalls from outside the guest: stock VMM, patched Firecracker with KVM guest-debug, host eBPF, sidecar monitoring, and Intel PT. The engineering reality is no such capture can be authoritative under a hostile guest kernel, where process identity and syscall-entry semantics become attacker-authored state. Introspection solves placement rather than attribution. Substrate alteration shifts the architecture along a curve, trading boundary strength for observation fidelity; gVisor detects syscalls because it implements them, thereby expanding the escape surface. Consequently, I signed the gap instead of concealing it. This architecture treats "unavailable, not observed" as a first-class signed state, causes observation loss to fail closed to indeterminate, and invalidates the run upon detected escape or observation-root failure. The pass condition for the adversarial suite is not the absence of a successful exploit, but that no exploit succeed silently. Attendees leave equipped to answer a fundamental telemetry question: what does data look like when the system lies?