Capability is advancing without a common scorecard
GAO's November 20 technology assessment identified four policy goals and 16 policy options for using AI in cybersecurity and defending against AI-enabled attacks. A central implementation problem is the absence of widely accepted evaluation criteria: buyers and regulators cannot compare reliability, adversarial resilience and failure behavior through one recognized test regime.
That gap matters because cyber tools act in environments where a false negative can miss an intrusion and a false positive can disrupt legitimate systems. Marketing claims about model accuracy do not substitute for testing against the organization's own network and threat model.
| Control | Accountability purpose |
|---|---|
| Human authorization | Gate consequential AI actions |
| Deterministic fallback | Maintain a non-AI recovery path |
| Pre-deployment validation | Test before operational use |
| Continuous evaluation | Detect changing performance and threats |
The threat side is no longer hypothetical
GAO cites a June 2026 AI-powered malware laboratory designed to help evade detection and a February incident in which a low-skill actor combined a commercial generative-AI tool with stolen credentials. These are examples in the federal assessment, not evidence that all attackers now possess autonomous cyber weapons.
AI can also help defenders triage alerts, analyze code and automate repetitive work. The accountability question is whether speed is paired with authority limits, audit logs and a safe path when the model behaves unpredictably.
Policy options are choices, not GAO endorsements
GAO explicitly presents trade-offs rather than a single mandate. Human review can slow response; shared benchmarks can become stale; extensive logging can add cost and create sensitive records. The report's value is to make those trade-offs visible before agencies or companies delegate consequential actions.
This analysis grouped GAO's 16 options around testability and operational control. It does not establish that any named agency deployed an unsafe system. The practical next record is a procurement requirement showing test datasets, pass/fail thresholds, human override and incident reporting.
