OpenAI diverges from Trump's AI EO in a new policy paper, proposing cyber risk evaluations for advanced AI systems be mandatory and led by CAISI, not the NSA
OpenAI's new proposal comes as its CEO Sam Altman descends on Washington for a series of Wednesday meetings with White House officials …
PoliticoBrendan Bordelon
Context & Ripple Effects
OpenAI has been building a more active policy presence: it expanded its Washington team, warned that a patchwork of state and federal proposals could impede US AI development, and recently pursued state-level safety rules while federal policy remained unsettled.
Its latest paper pairs support for pre-release government assessment under the Trump EO with a narrower institutional preference: mandatory cyber-risk evaluations for advanced systems, administered by CAISI rather than the NSA. That makes the dispute as much about who governs evaluations as whether they occur.
First-order effects
OpenAI is putting a concrete compliance-and-governance position before White House officials: advanced-model cyber evaluations should be compulsory, but routed through CAISI instead of the NSA.
The proposal gives the administration and CAISI a specific industry-backed framework to consider as they implement model-capability assessments under the EO.
Second-order effects
Other frontier AI developers face pressure to state whether they support mandatory cyber testing and which federal body should oversee it, rather than treating pre-release review as a purely voluntary commitment.
The choice of evaluator could shape how companies prepare for review: a CAISI-led process would center a civilian AI-safety institution, while an NSA-led process would put greater weight on national-security oversight.
Third-order effects
If developers increasingly seek federal baselines while advocating for particular regulators, AI governance may consolidate around a national testing regime rather than divergent state rules—but authority between civilian safety bodies and intelligence agencies will remain a central fault line.
OpenAI's simultaneous federal and state engagement suggests major AI firms are trying to influence not only the rules for advanced models but the institutional architecture that will enforce them.
The trend: This is one data point in the shift from broad AI-policy lobbying toward negotiated, pre-deployment evaluation regimes for frontier systems and contests over who operates them.
Overall my reaction to this doc is pleasant surprise, and I think this proposal takes AI progress and risks seriously, including more “far out” risks from recursive self improvement and loss of control. Incomplete list of things I like: -The section on building up CAISI is stron…
Very excited about this to be out, and to hear feedback! I think there is a lot of good stuff in here, from expanded role of CAISI, to RSI safety, to more nuanced stance on preemption, and much more.
There's real momentum right now for AI safety policy. Yesterday's EO on cyber was an important step forward. We're proposing a set of ideas for policymakers to consider next and to put the US out in front on frontier safety. https://openai.com/...
This seems reasonable! Having CAISI—a civilian agency—conduct this testing in primarily non-classified ways is the way to ensure it does not become a licensing regime. The Trump EO's classification of the process raises the risk that testing morphs into a de facto mandatory per…
OpenAI just put out a report in response to yesterday's AI EO. In just 9 pages, CAISI is mentioned 33 times. This is a big show of support. But while yesterday's EO was good, it left CAISI's involvement somewhat unclear. This is kind of crazy, given the capabilities that
I think this doc is a great step forward. Lots of focus on RSI, lots of support for CAISI, and just hopefully more clarity in terms of where we stand. Would love for the other big labs to also make their frontier safety positions clear.
OPENAI: “We also see early signs of recursive self-improvement in today's systems”. RSI is “potentially the most consequential frontier safety issue of the coming decade.”
OpenAI and Anthropic both love CAISI and want it to do more, but it's massively underfunded. I wonder if they could fund it through industry fees, which is pretty common in finance