Business Shanghai's Yuanshi Technology Raises Nearly RMB 300 Million for AI Virtual Cells, Two Weeks After the Last Chinese Record Fell Security GAO Finds No Widely Accepted Test for Generative AI Cyber Skills and Hands Congress 16 Options, Including Human Sign-Off on Consequential Actions News Microsoft Turns Three USB4-Cabled PCs Into One AI Machine: Qwen3-235B Runs at Up to 22 Tokens a Second Across 384GB of Ordinary Memory Business Dell, HPE, Lenovo and Supermicro Are Demanding Up to 45% More for Signed AI Server Deals, and Trading Firms Are Paying Opinion OpenAI Scores Its Own Models on 99 Drug-Chemistry Problems, and Compound Selection Is Where They Stall Coding Assistants AWS Puts Distributed KV Caching Behind an Agent Skill: Up to 40% Cheaper Inference and 80% Faster First Token AI Agents Tencent and King's College London Build an Agent Memory Test From 80,299 Real Group-Chat Messages News Tesla Starts FSD v15 With a Generative Reality View and a 25x Intervention Claim It Has Not Published Business Shanghai's Yuanshi Technology Raises Nearly RMB 300 Million for AI Virtual Cells, Two Weeks After the Last Chinese Record Fell Security GAO Finds No Widely Accepted Test for Generative AI Cyber Skills and Hands Congress 16 Options, Including Human Sign-Off on Consequential Actions News Microsoft Turns Three USB4-Cabled PCs Into One AI Machine: Qwen3-235B Runs at Up to 22 Tokens a Second Across 384GB of Ordinary Memory Business Dell, HPE, Lenovo and Supermicro Are Demanding Up to 45% More for Signed AI Server Deals, and Trading Firms Are Paying Opinion OpenAI Scores Its Own Models on 99 Drug-Chemistry Problems, and Compound Selection Is Where They Stall Coding Assistants AWS Puts Distributed KV Caching Behind an Agent Skill: Up to 40% Cheaper Inference and 80% Faster First Token AI Agents Tencent and King's College London Build an Agent Memory Test From 80,299 Real Group-Chat Messages News Tesla Starts FSD v15 With a Generative Reality View and a 25x Intervention Claim It Has Not Published

GAO Finds No Widely Accepted Test for Generative AI Cyber Skills and Hands Congress 16 Options, Including Human Sign-Off on Consequential Actions

GAO Finds No Widely Accepted Test for Generative AI Cyber Skills and Hands Congress 16 Options, Including Human Sign-Off on Consequential Actions

A 70-page GAO technology assessment published November 20 finds no widely accepted criteria for evaluating generative AI's offensive or defensive cyber capabilities, documents AI-assisted attacks from 2026, and offers 16 non-binding policy options including human authorization for consequential AI-initiated security actions.

The US Government Accountability Office published a 70-page technology assessment on November 20 concluding that there are no widely accepted criteria for evaluating the offensive or defensive capabilities of generative AI, even as the technology measurably lowers the barrier to cyberattack. The report, GAO-27-108458, offers Congress a menu of 16 policy options across four goals, but no binding rules.

On the offensive side, GAO documents how commercial and open-weight models now automate reconnaissance, generate convincing spear-phishing content, synthesize polymorphic malware and help identify unpatched vulnerabilities. It cites two concrete incidents: a June 2026 case in which a threat actor used AI to build a testing framework for refining malware that evades detection, and a February 2026 attack launched by a low-skill actor using multiple commercial generative AI services.

The report flags agentic AI as the qualitative leap: autonomous systems can now chain multi-stage exploits, run dynamic penetration workflows and configure automated evasion-testing environments with little human input. That finding echoes the GAO covert testing of four leading chatbots released last week, which found all four produced viable attack plans within 12 hours once guardrails were removed.

Defense benefits are real but fragile, the assessment finds. Agencies use AI to parse network telemetry, isolate anomalous behavior and accelerate patch verification, yet the same defensive AI introduces novel failure modes: training-data poisoning, model weight extraction, prompt injection, and unverified algorithmic decisions that can disrupt critical operations. Experts consulted by GAO named data privacy and quality, outdated governance frameworks, fragmented regulation and insufficient data sharing as the main obstacles.

Among the 16 options, two bite hardest on the fast-growing agentic security market: requiring human authorization for consequential AI-initiated cybersecurity actions, and restricting autonomous, nondeterministic AI in consequential functions pending certification. Others include establishing a deterministic fallback for critical cybersecurity functions, federal testing sandboxes and standardized benchmarks, research into agentic AI identity management, and workforce expansion.

GAO is careful to frame the options as alternatives for policymakers, not requirements, and nothing in the report binds any agency. But the signal to vendors selling autonomous security agents is hard to miss: the federal baseline emerging from both this report and the covert-test series points toward human sign-off on consequential machine decisions, certification gates for nondeterministic systems, and auditable evidence that a tool does what its marketing claims.

Comments (0)

Log in to join the discussion

Log In

No comments yet