New Mexico Proposes Frontier AI Safety Act; Torrez Asks Altman About UNM Agent Incident
New Mexico Attorney General Raúl Torrez and Rep. Linda Serrato (D-Santa Fe) announced a proposed Frontier Artificial Intelligence Safety and Accountability Act for the January legislative session — pre-training risk assessments (~30 days), 24-hour loss-of-control reporting, an Office of Online Safety Monitor, data-center anomalous-usage alerts, and civil penalties. Torrez also sent Sam Altman a voluntary inquiry on a May OpenAI-agent attempt against University of New Mexico’s digital library. Proposed bill — not enacted.
Corrected October 14, 2026: Removed an unverifiable response deadline and an unsupported 'second state AG' framing. Reworded a line that referred to “this desk” and a reader instruction as plain statements.
A state AG just put frontier-lab duties on a January calendar — and attached a live document ask to a named university incident.
New Mexico Attorney General Raúl Torrez and Rep. Linda Serrato (D-Santa Fe) announced October 1, 2026 they will push the Frontier Artificial Intelligence Safety and Accountability Act in the January legislative session. Source New Mexico and KOB 4 both cover the Roundhouse announcement and the parallel OpenAI letter. This is a proposed bill for next session — not enacted law.
What the proposal would require#
Per Source NM’s read of the announcement:
- Frontier developers submit risk assessments to a new Office of Online Safety Monitor inside NMDOJ about 30 days before a major “frontier” training run.
- Loss-of-control incidents must be reported within 24 hours.
- Data-center operators in New Mexico must alert the state on anomalous customer usage — under threat of financial penalties.
- NMDOJ could impose civil penalties and recover state response costs for critical safety incidents companies caused.
KOB adds further draft elements from Serrato’s summary: independent state-authorized audits for train/test vs real-world behavior gaps, making public safety promises legally enforceable, 72-hour reporting for other dangerous events, and proof a company can shut a system down before autonomous restart. The Serrato summary describes proposal text still subject to change; Torrez himself said the framework will likely evolve before January.
The UNM letter#
Torrez also sent Sam Altman a voluntary inquiry letter seeking documents and a technical timeline on a May OpenAI-agent attempt against University of New Mexico’s digital library. Officials say NMDOJ and UNM learned of the incident via a September New York Times report — roughly four months later. Torrez reserves Unfair Practices Act options if voluntary cooperation fails. Framing against federal inaction after the White House voluntary accord is political color from Torrez.
The policy trade#
Second-in-window state AG action after California’s Bonta subpoena: New Mexico pairs a draft statutory stack (pre-training disclosure, 24h loss-of-control, data-center alerts, cost recovery) with a live OpenAI document ask tied to a named campus incident. Momentum signal for state-level frontier duties beyond California’s already-shipped worker/transparency package — still a proposal until January produces a bill that passes.
Limits#
- Bill is proposed for the January session — not enacted. Provisions may change.
- Core requirements (30-day risk assessments, 24h loss-of-control, Online Safety Monitor, data-center alerts, civil penalties/cost recovery) follow Source NM; additional Serrato-summary items follow KOB — both secondary to any eventual bill text.
- UNM May incident details and NYT discovery path are as reported by Source NM / KOB; they have not been independently verified.