The US government issued an export control directive, suspending access to Anthropic's Fable 5 and Mythos 5 AI models for foreign nationals due to national security concerns, leading Anthropic to disable access for all users.
The speaker criticizes Anthropic for its "gaslighting" marketing, having previously claimed its models were too powerful to release, and for pushing for government regulation of AI while refusing to address a "jailbreak" vulnerability in Fable.
Anthropic's billing practices are also scrutinized, as they appear to claw back functionality from subscription plans, forcing users to more expensive API-metered rates while framing it as a "credit."
The government's directive is deemed impractical and counterproductive, as it necessitates identity verification for online services and bars foreign-born AI researchers—who constitute a significant portion of US AI talent—from accessing and thus fixing the models.
Both Anthropic and the government are accused of hypocrisy and irrationality, with the government's safety claims undermined by allowing US citizens access to potentially dangerous AI, and Anthropic's safety branding contradicted by its actions.
Anthropic's official statement regarding the US government's directive to suspend access to Fable 5 and Mythos 5 models.
[
00:05:19
]
Government Restriction on Anthropic's AI Models [00:00:18]
US Government Issues Export Control Directive [00:00:18]
Anthropic's newest and most powerful models, Fable 5 and Mythos 5, were restricted by the US government. [00:00:22]
The directive cited national security authorities, prohibiting foreign nationals from accessing these models, whether inside or outside the United States. [00:05:17]
In response, Anthropic fully disabled access to both models for all customers to ensure compliance, stating that access to other Anthropic models would not be affected. [00:05:32]
The speaker questions how a company can effectively restrict foreign nationals without implementing burdensome identity verification for all users, such as requiring passports or birth certificates, which would fundamentally alter internet access. [00:05:40]
He argues that such restrictions are easily circumvented (e.g., sharing login credentials with a US friend) and are ultimately ineffective, drawing parallels to failed attempts to control media sharing online. [00:06:57]
Anthropic's Contradictory Stance on AI Safety and Regulation [00:01:30]
Anthropic previously marketed their models as "too powerful" and "too dangerous" to release, creating a narrative of advanced capability that the speaker views as "gaslighting." [00:01:33]
The company publicly advocated for stringent government regulation of AI, proposing mandatory third-party testing for models above a certain computational threshold. [00:08:05]
They also suggested the government should have the power to block or deter deployment of models deemed unsafe. [00:08:57]
This advocacy came despite a prior dispute where the Trump administration declared Anthropic a "supply chain risk" for refusing to allow its models to be used for mass surveillance or fully autonomous weapons by the Department of Defense. [00:04:23]
The speaker finds it ironic that Anthropic pushed for government oversight, then expressed surprise and disagreement when the government applied such controls to their own models. [00:09:12]
A screenshot from Anthropic's statement details their proposal for mandatory testing and government authority to block AI models based on risk assessment.
[
00:08:48
]
The "Jailbreak" and Anthropic's Response [00:13:32]
A trusted partner and the US government, while testing Fable, discovered a "jailbreak" method that bypassed its guardrails, potentially exposing its advanced cyber capabilities. [00:14:06]
The administration requested Anthropic fix the vulnerability or redeploy the model. [00:14:13]
Anthropic, specifically CEO Dario, reportedly refused, downplaying the seriousness of the jailbreak in a blog post, calling it "simple minor vulnerabilities" and claiming "no testers have been able to find a universal jailbreak." [00:14:17]
The speaker interprets Anthropic's denial as obfuscation, suggesting a jailbreak was indeed found but downplayed. [00:16:36]
This refusal led to the government's export control directive. [00:14:45]
The government expressed bewilderment at Anthropic's unwillingness to cooperate, given the company's prior emphasis on safety. [00:15:09]
Anthropic's Billing Practices: A "Drug Dealer" Analogy [00:10:03]
The speaker criticizes Anthropic's shift in billing for programmatic usage of its Claude models. [00:10:03]
Initially, users could use Claude programmatically via SDKs or scripts as part of their subscription. [00:10:08]
Anthropic announced a "dedicated monthly credit for programmatic usage" [00:10:15].
However, this actually means that programmatic usage, which was previously included in subscription plans, now moves to an API-metered billing structure. [00:10:40]
API rates are significantly more expensive (20-30 times higher) than the fixed monthly subscription cost. [00:12:15]
The speaker compares this tactic to drug dealers giving free samples and then jacking up prices once users are hooked, calling it "gaslighting" and dishonest. [00:11:23]
A tweet from ClaudeDevs announces a "dedicated monthly credit for programmatic usage" for paid Claude plans.
[
00:10:17
]
A clarification tweet reveals a diagram illustrating separate pools for interactive and programmatic usage, with programmatic usage now metered at API rates.
[
00:12:06
]
The speaker highlights a significant flaw in the government's directive: 59% of top-tier AI researchers in US institutions are foreign-born. [00:17:52]
This means the very individuals most capable of fixing security vulnerabilities in AI models are now legally barred from accessing and working on them. [00:18:02]
This creates an impossible situation for Anthropic to comply with the government's demands while maintaining its workforce. [00:18:15]
The speaker argues that if the models are truly as dangerous as claimed (e.g., capable of hacking banks or government infrastructure), restricting them only to foreign nationals makes no sense. [00:19:04]
If safety is the primary concern, nobody, including US citizens, should have access to such powerful models, as domestic actors could also exploit them. [00:19:46]
Both the government and Anthropic are criticized for their "silly" and hypocritical posturing, where claims of safety and responsibility mask other agendas (e.g., control, profit). [00:19:15]
