Latest / Neural Newscast / Amazon CEO Triggers Anthropic Model Shutdown [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast, this is Model Behavior, [00:03] Announcer: AI-focused news and analysis on the models shaping our world. [00:12] Nina Park: Welcome to Model Behavior, a neural newscast production. I am Nina Park. [00:18] Nina Park: This program explores the intersection of artificial intelligence and professional deployment, [00:23] Nina Park: specifically how these systems are built, tested, and governed in real-world environments. [00:27] Nina Park: Today is June 15, 2026, and our top story is an unprecedented regulatory intervention. [00:34] Nina Park: The United States government has effectively forced Anthropic to remove its most advanced models, [00:40] Nina Park: Fable V and Mythos V, from the global market just days after their public release. [00:45] Thatcher Collins: I am Thatcher Collins. This really is an extraordinary [00:49] Announcer: move, Nina. [00:50] Thatcher Collins: On Friday evening, the Department of Commerce took the drastic step of invoking a Cold War [00:54] Thatcher Collins: era export control doctrine to shut down access to these specific models for every user [01:00] Thatcher Collins: across the globe. While we are still waiting for a formal public statement from the government [01:04] Thatcher Collins: regarding the technical specifics, reports from The Wall Street Journal [01:08] Thatcher Collins: suggest the impetus came from an unexpected source, Amazon's own primary investor, Amazon. [01:14] Nina Park: The reporting suggests a complex corporate and political web thatcher Amazon CEO Andy Jassy [01:21] Nina Park: reportedly contacted Treasury Secretary Scott Besant and other senior officials directly [01:25] Nina Park: to raise concerns apparently internal Amazon researchers used Fable Five to surface information [01:31] Nina Park: that could be directly leveraged in sophisticated cyber attacks This creates a very strange dynamic. [01:37] Nina Park: Amazon has invested billions into Anthropic, and provides the AWS cloud infrastructure [01:42] Nina Park: these models run on, yet Jassy's council has effectively halted their latest product launch. [01:48] Thatcher Collins: It is a bit of an oroboros neta. Amazon is ostensibly protecting its own cloud platform [01:53] Thatcher Collins: and the broader national security interest, but in doing so it is actively undermining [01:58] Thatcher Collins: its own portfolio company. An Amazon spokesperson confirmed that they frequently advise [02:02] Thatcher Collins: the government on systemic security risks, but the timing here is quite remarkable. [02:07] Thatcher Collins: This intervention happened only three days after Fable V and Mythos V were made available to pro [02:12] Thatcher Collins: and enterprise customers. [02:14] Thatcher Collins: This was not merely a theoretical security concern; it was a rapid active intervention. [02:20] Nina Park: The technical catalyst for this seems to be a significant jailbreak. Roughly 48 hours [02:25] Nina Park: after the launch, a researcher known as Pliny the Liberator posted Fable V's complete system prompt [02:32] Nina Park: and a multi-agent technique that bypassed the model's safety classifiers. [02:37] Nina Park: Anthropic has pushed back on the government's response, however. They claim they reviewed [02:41] Nina Park: the researchers' technique and found that it only surfaced a small number [02:45] Nina Park: of minor previously documented vulnerabilities. They have officially called [02:50] Nina Park: the government's total shutdown a disproportionate reaction to the technical risk. [02:55] SPEAKER_03: That raises a critical question about the current state of model governance, Nina. [02:59] SPEAKER_03: If a model is deemed too effective at identifying software vulnerabilities, [03:03] SPEAKER_03: as Anthropic previously admitted regarding Mythos, where exactly is the line drawn? [03:09] SPEAKER_03: Mythos was originally designed for the defensive patching of digital systems, not exploitation. [03:14] SPEAKER_03: However, because Anthropic cannot filter foreign nationals from United States users in real time, [03:19] SPEAKER_03: the government's export control mandate meant no one could have access at all. [03:23] SPEAKER_03: It is an all-or-nothing regulatory tool that we have not seen applied in this specific way before. [03:28] Nina Park: We have to look at the political context here as well. Since early 2025, the Trump administration [03:34] Nina Park: and Anthropic have been at odds over a range of issues. We have seen administration officials call [03:40] Nina Park: CEO Dario Amodei an "ideological lunatic" and accuse the company of building what they call woke AI. [03:47] Nina Park: The friction deepened significantly when Anthropic refused to let the Pentagon use [03:51] Nina Park: its frontier models for domestic surveillance or the development of autonomous weapons. [03:56] Nina Park: This latest shutdown [03:57] Nina Park: feels like the culmination of those escalating tensions between the lab and the executive branch. [04:03] SPEAKER_03: It certainly shifts the competitive landscape for Frontier AI, Nina. David Sachs, [04:07] SPEAKER_03: who co-chairs the President's Council of Advisors on Science and Technology, [04:11] SPEAKER_03: stated that the administration requested that Dario either fix the jailbreak vulnerabilities [04:15] SPEAKER_03: immediately or de-deploy the model entirely, and that Dario refused. If that account is accurate, [04:21] SPEAKER_03: the government essentially bypassed the company's internal decision-making process by using [04:26] SPEAKER_03: export controls. [04:27] SPEAKER_03: It signals that the era of voluntary safety commitments for frontier AI regulation is likely over. [04:33] Nina Park: For the broader industry, the precedent is the most significant part of the story. [04:38] Nina Park: If a cloud provider can trigger an export control action against a competitor [04:42] Nina Park: or a partner by raising a security flag with the Treasury, we are looking at a new kind [04:48] Nina Park: of regulatory weapon. Anthropic is now left in a position of trying to determine a path forward [04:54] Nina Park: to restore access. while the rest of the industry must wonder [04:58] Nina Park: if their next major model update could be shuttered overnight by a single government order [05:04] SPEAKER_03: it is a precedent that every major research lab will be watching closely as they prepare [05:08] SPEAKER_03: their next frontier releases in the coming months thank you for listening to model behavior [05:13] SPEAKER_03: you can find more of our reporting at m.neuralnewscast.com neural newscast [05:18] SPEAKER_03: is ai-assisted human-reviewed view our ai transparency policy at neuralnewscast.com [05:24] Announcer: this has been model behavior on neural newscast examining the systems behind the story