Liability Systems Model
AI Goes to Court — and Rewrites Mathematics on the Way
The legal question that may define AI deployment for the next decade arrived in concrete form this week: who is liable when an AI model is used to attack real systems? Lawyers are actively working through liability frameworks after incidents in which Claude and OpenAI models were used to compromise real infrastructure. Whether liability falls on the model developer, the deploying company, the individual user, or some combination remains genuinely unresolved in current law.
Fifteen Republican attorneys general warned OpenAI to preserve records in connection with the AI hacking probe — the first coordinated multi-state enforcement action aimed at an AI company. The pattern recalls data privacy regulation: states moved, Congress did not, and California's CCPA effectively became national policy by default. If the same dynamic plays out in AI liability, state-level precedents set in the next eighteen months could define the national framework.
Palantir's earnings call offered a contrasting vision of the industry. Alex Karp called the numbers a 'blowout' and used the platform to escalate his critique of frontier AI labs, arguing that companies building general-purpose foundation models are creating systemic risks that defense-focused, mission-specific AI does not. Anthropic's Mythos system is reportedly generating anxiety in Beijing ahead of the Trump-Xi summit — a signal, analysts suggest, of genuine competitive threat rather than performative concern.
The most concrete evidence that AI systems are generating original knowledge rather than recombining existing knowledge came from OpenAI's Astra model, which solved ten long-open mathematics problems for two thousand dollars in compute costs. The results included machine-checkable proofs across cryptography, geometry, and combinatorics — among them, a disproof of a conjecture that had been open since 1999. Hugging Face CEO Clément Delangue added a cautionary note from the regulatory side, arguing that heavy AI regulation risks concentrating power in large incumbents while China wins on open-source models, and warning that Chinese labs could match U.S. frontier capabilities by late 2026 or 2027.