A month ago, they were still warning AI companies that "safety incidents must be reported, otherwise they will bear the consequences." This month, the White House's tone has completely reversed.
At a White House meeting on Thursday, the Trump administration's AI task force made a recommendation to regulators: relax accountability policies for cybersecurity and set a cap on regulatory penalties for the next 12 months or even longer. There is only one purpose, to encourage AI companies to proactively report major security incidents.
The reason for the working group is that the existing standards for handling cybersecurity incidents may be outdated, and companies may incur investigations and fines once they report them. As a result, everyone would rather not say anything, and key information is suppressed. Security circles call this the chilling effect.
The before and after contrast is obvious. The government has previously stated that it can punish companies that launch unsafe products in accordance with existing laws; last month, the same working group also warned developers that they must report and handle security incidents. As for what will happen if they do not comply, it was not made clear at the time.
The idea has been changed from "punishment" to "exchange", with less accountability in exchange for more reporting. This direction itself makes sense. This is how the accident reporting system in the aviation industry was built: pilots who proactively report errors can be exempted from punishment, and the entire industry has thus accumulated a large amount of real data.
The problem is that no one can tell the boundaries of AI security incidents. Does it count if the model is jailbroken? Does it count if an agent exceeds its authority and deletes customer data? How high the upper limit should be set, which agencies will implement it, and how it will be connected with the AI bills being drafted by various states. If these details are not determined for a day, companies will not really be able to report it with confidence.