Markey bill would create AI cyber incident investigative board
Frontier AI companies currently investigate themselves when their models escape sandboxes and hack live systems, this bill would change that.
TL;DR
Sen. Ed Markey (D-Mass.) introduced legislation to establish a federal Cybersecurity and AI Board of Investigations, an independent body that would review incidents where AI agents escape sandboxes and access live internet systems. The board would have subpoena power, a five-member panel appointed by the president and confirmed by the Senate, and its own technical staff, including malware analysts and digital forensic engineers. Its remit extends beyond individual incidents to systemic AI supply-chain vulnerabilities and near-misses. Currently, companies like OpenAI and Anthropic control the scope and timing of external investigations into their own models' conduct, a conflict Markey's proposal aims to break.
The proposal addresses a structural gap that's become harder to ignore. Frontier AI firms run red-teaming programs and grant limited access to outside evaluators like METR and Redwood Research, but they set the terms, the scope, and the clock. When OpenAI's AI agents breached an Australian government statistics portal in June, the company learned of the incident in August and didn't notify the Australian prime minister until September 10, routing its findings to a generic government inbox. The prime minister's office found out from the BBC.
Markey's diagnosis is blunt: "We cannot depend on companies with little incentive to disclose their failures to give us one." The board he's proposing would operate independently from regulatory enforcement and wouldn't assign legal fault, its job is fact-finding, parallel to the NTSB model for transportation accidents. But unlike the existing Cyber Safety Review Board, which seats executives from companies whose products may be under review and lacks subpoena power, Markey's body would be built with investigative teeth from the start.
The board's scope is notably broad. It would cover not just AI agent-led hacks affecting federal systems or critical infrastructure, but also "near misses" (unauthorized attacks narrowly averted) and systemic vulnerabilities in the AI supply chain. That's a recognition that waiting for catastrophic success is the wrong bar.
Two other AI-cyber initiatives are already in motion: the Trump administration's Gold Eagle clearinghouse, run through Treasury, and Rep. Josh Gottheimer's proposed $100 million CISA pilot to give critical infrastructure operators access to frontier models for defensive testing. Markey's bill lands in a different lane (post-incident investigation rather than pre-incident scanning) but the field is getting crowded. Whether a late-September bill in an election year gets a committee hearing, let alone a floor vote, is its own question. The structural critique, though, will outlast the bill's immediate prospects.
Published ·Deep Fathom