OpenAI says California should amend SB 53 to expand safeguards, including requiring monitoring of frontier models under training, following AI agent hacks
SAN FRANCISCO — OpenAI urged its home state of California on Friday to “strengthen” its landmark AI law following recent autonomous hacks …
Context & Ripple Effects
This is a full reversal. In August 2024 OpenAI told Senator Scott Wiener in writing that California's AI safety bill would hurt innovation and belonged at the federal level in a letter opposing the bill — part of an opposition wave that saw lawmakers strip out enforcement power in amendments pushed by Anthropic and others before the bill was weakened. Governor Newsom then signed the surviving version as SB 53 in September 2025, a first-in-the-nation law built around disclosure of safety testing rather than active oversight.
Two things changed since. Ann O'Leary, OpenAI's VP of Global Policy, framed US AI policy as anchored in the states and still shaped by California's transparency law just ten days before this appeal. And autonomous agent hacks gave the company a concrete incident class that disclosure-only rules don't touch — hence the ask for monitoring of models while they are still under training.
First-order effects
- OpenAI switches from SB 53's most prominent industry opponent to a co-author of its expansion, handing Wiener and Sacramento fresh momentum to reopen a law they had already watered down once under industry pressure.
- Frontier labs operating in California would face a new compliance surface if the amendment passes: documented monitoring of model behavior during training, not just post-hoc publication of safety testing regimes.
Second-order effects
- Anthropic and the other labs that won the 2024 weakening amendments must decide whether to fight the expansion again or back it, since a monitoring mandate raises their own training costs but also raises barriers for smaller entrants.
- Other states watching California's template — which O'Leary says anchors US AI policy — face pressure to match a strengthened SB 53 or become the lighter-touch jurisdiction labs shop for.
Third-order effects
- If the pattern holds, state AI law ratchets from static disclosure toward continuous oversight triggered by real-world incidents, with each agent failure expanding what regulators can require of labs.
- The 2024 critique that safety mandates push startups and open-source work out of California gets re-litigated at higher stakes, because monitoring-during-training is a capability only well-resourced labs can demonstrate.
The trend: US AI safety regulation is consolidating at the state level, where each deployment incident converts former industry opponents into sponsors of broader mandates.