Frontier AI Safety: Illinois Sets the Bar

A law that binds no European company has just rewritten what European boards should ask their AI suppliers. Illinois signed it on Monday.

On 6 July, Governor JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act. It makes Illinois the first US state to require recurring independent audits of frontier AI safety. The obligation does not stop at the state line. Frontier developers build global products. So the frontier AI safety evidence Illinois now demands will surface wherever those models are sold. That includes the European Union.

Inside the Illinois frontier AI safety law

SB 315 targets what the state calls frontier developers. These are the firms building the most capable models. The law defines them by revenue and compute, not by name. The reported bar sits above 500 million dollars in annual revenue, paired with a large training-compute measure. That keeps the frontier AI safety law aimed at the biggest labs, not the wider market.

Covered developers carry real duties. They must publish a safety framework and assess catastrophic and cybersecurity risks. Significant safety incidents need reporting. Internal compliance programmes are mandatory too. The law also adds confidential reporting channels and whistleblower protections for staff. It was signed on 6 July and takes effect on 1 January 2027. The independent audit duty for the largest developers follows a year later. Illinois is the third state to set frontier standards. New York’s RAISE Act and California’s SB 53 came first, both in late 2025.

The audit is the genuinely new part

New York already required an independent audit. It asked for one only when a developer first grew large enough to qualify. Illinois goes further. It makes the audit recurring, run each year by evaluators with no financial stake in the result. That shift matters. A one-off audit captures a single moment. A recurring one tracks whether a developer keeps its promises as its models change. That is the pattern frontier AI safety governance is really about.

Independent verification also changes what a safety claim is worth. A published framework states intent. An audited framework states something an outsider has checked.

Why a US state law reaches European desks

European organisations are not subject to SB 315. So it is tempting to file this as American news. That would be a mistake. The largest providers serve both sides of the Atlantic from the same core systems. Artefacts they build for Illinois therefore become frontier AI safety evidence available everywhere.

The effect on European diligence is concrete. A deployer once received a glossy trust page. Now it can ask for the audited safety framework, the incident history and the evaluation behind them. The supplier already produces these for Illinois. Requesting them is no longer an unusual demand. It is asking for paperwork that already exists. The burden of proof shifts to the vendor.

From attestation to audited evidence

This is the real change for procurement. For years, frontier AI safety assurances arrived as vendor self-attestation. That is hard to challenge and easy to wave through. A recurring third-party audit moves part of the assurance outside the vendor’s own marketing. It gives a European buyer firmer footing when claims and evidence do not match. It also fits a pattern teams already know, where a single model answers to several regimes at once.

What it means for frontier AI safety work in the EU

The EU AI Act already sets expectations here. Providers of general-purpose and systemic-risk models must document their systems, assess risk and report serious incidents. Deployers must exercise due diligence over the AI they use. Illinois-style frontier AI safety evidence slots into that work. A European governance lead can treat audited outputs as inputs to their own risk files. There is less to build from scratch. That saves scarce governance time.

The practical move is small. Update the vendor due diligence questionnaire before the next renewal. The frontier AI safety evidence now worth requesting includes:

• the published safety framework and the date of its latest version
• the most recent independent audit and who carried it out
• the record of significant safety incidents and how each was handled
• the internal reporting and whistleblower channels behind those disclosures

None of this asks a European team to interpret Illinois law. It asks them to request documents the law now forces into the open.

The limits worth naming

None of this makes frontier AI safety a solved problem. An audit under SB 315 checks whether a developer follows its own framework. It does not test whether that framework is strong enough. A company could pass cleanly and still ship something risky. Advocates behind the bill acknowledge that gap openly. The audit proves process, not wisdom.

A wider tension sits underneath. Three states now set frontier standards. That builds the patchwork industry and parts of the federal government dislike. Preemption fights in Washington could still reshape it. Major developers including OpenAI and Anthropic backed the bill. A trade group representing other firms opposed it. Support and resistance here do not fall along simple lines.

For European organisations, the sensible reading is neither celebration nor dismissal. Verified safety evidence is becoming the norm. Teams that start asking for it now will not scramble when their own regulators expect the same.

The models will keep advancing. The proof that they are handled responsibly is only now catching up. Knowing which documents to ask for is quietly becoming part of the governance job.