At DEF CON, hobbyists and academics circumvented the safety guardrails of frontier AI models, prompting them to generate instructions for illicit surveillance and disinformation. Meanwhile, the U.S. government secured pre-release access to these same foundation models.
Despite White House agreements, the government's AI safety evaluation protocols remain a black box. The U.S. government, now the world’s primary AI safety gatekeeper, offers no public insight into its benchmarks, testing methodologies, or the consequences for failing an evaluation. This opacity forces the public and independent researchers to trust the government's undisclosed safety standards, leaving no room to verify whether these powerful systems are truly safe for widespread deployment.
NIST Director Laurie E. Locascio states the official goal is to build "a culture of AI safety and responsibility"; however, this governance framework is being drafted behind closed doors.
The "Voluntary" Agreements
The White House frames these pre-deployment safety evaluations as "voluntary commitments". However, this voluntarism operates within a compulsory framework established by President Biden's Executive Order 14110. The order mandates that government evaluations prioritize specific national security risks, including threats related to chemical, biological, radiological, and nuclear (CBRN) capabilities.
While leading developers from Anthropic to Google and Microsoft have agreed to participate, they are effectively submitting to a mandatory, national security-focused risk assessment. The U.S. AI Safety Institute (USAISI) is not asking companies what they think is dangerous; it is evaluating these models against a specific threat model defined by the White House. For developers and the public, this narrow focus on catastrophic CBRN risks means that other significant harms—such as algorithmic bias, economic disruption, or the proliferation of sophisticated surveillance tools—may be systematically downplayed or overlooked in the official safety assessment.
Sources & References
Scott Wolfe
In-depth analysis of trending global events
More from this author
Comments
Post a Comment