AI safety got a compute price this week, and AI liability got an exclusion

Two things landed within 48 hours of each other, and read together they move real money onto the enterprise side of the ledger.

On August 19 OpenAI said its largest planned frontier reinforcement learning run remains on hold(opens in a new tab). The company had already determined, on August 7, that its unreleased Astra model meets the critical cybersecurity threshold in its Preparedness Framework, defined as finding and developing working zero-day exploits against hardened real-world systems without human direction. The response includes activation classifiers that inspect every sampled token on Astra inference involving tools, escalating to automated investigators against a 30 minute alert target. OpenAI puts the cost of that monitoring at roughly 20 percent of the inference compute being monitored.

That number is the part worth writing down. Frontier safety controls have usually been described to buyers as a governance posture. This is the first widely reported figure that prices one as a compute line item, and a 20 percent overhead on monitored inference is not a rounding error in a unit economics model. Any capability that arrives with a critical classification attached will arrive carrying that kind of tax, and it will show up in vendor pricing rather than in a security budget.

The second event runs the other way. Insurance Journal reported on August 17 that carrier interest in AI exclusions is growing(opens in a new tab), with filings to adopt ISO endorsements CG 40 47, CG 40 48, and CG 35 08. Those forms took effect in January 2026 and exclude bodily injury, property damage, and personal and advertising injury arising out of or attributable to generative AI. Berkley has gone further with an absolute AI exclusion across specialty liability lines, written broadly enough that coverage counsel read it as reaching the use, deployment, integration, or failure of AI, not only AI-generated output. A company that merely buys AI tools is inside that language.

So the cost of containing frontier risk is becoming explicit and billable, while the coverage that would have absorbed the resulting loss is being withdrawn from general liability, D&O, and E&O. Enterprises are being asked to pay more for safety and to self-insure more of what safety fails to catch. Neither move was announced as a price change. Both are.