Anthropic Fable Shutdown Explained: National Security & Export Controls 2024

Anthropic Fable shutdown marks a pivotal moment in the intersection of artificial intelligence development and national security policy. On Friday night, following a series of escalating tensions, the government imposed stringent export controls that forced Anthropic to take the Fable 5 and Mythos 5 models entirely offline. This decision followed a directive from the U.S. Department of Commerce and revelations that the models were accessible to foreign nationals with ties to the Chinese Communist Party. The incident has raised critical questions about the limits of private sector autonomy in technology, the efficacy of current safety guardrails, and the growing friction between AI developers and the administration.

What is the Anthropic Fable Shutdown?

The Anthropic Fable shutdown refers to the immediate cessation of access to Anthropic's Fable 5 and Mythos 5 large language models. This action was taken in direct response to a letter sent by the U.S. Department of Commerce, which directed Anthropic to suspend all access to these models by any foreign national. The shutdown was precipitated by a convergence of factors, including the discovery of significant security vulnerabilities, concerns over jailbreak capabilities, and a breakdown in communication between the company and government officials.

The event underscores a shift in how the United States approaches AI governance. An administration official stated that Anthropic failed to honor a recent cyber executive order, leading to the threat of export controls. Furthermore, officials indicated that Anthropic knew a jailbreak could happen but chose to distribute the model anyway, prioritizing release over the specific safety protocols demanded by the government. The shutdown effectively removed these 'Mythos-class' models from public and partner use, signaling a new era where government oversight can instantly override commercial deployment strategies.

Timeline: From Mythos Release to Model Take-Down

The sequence of events leading to the Anthropic Fable shutdown unfolded rapidly over a span of several months, moving from initial friction to a decisive government intervention.

The timeline begins in February and March, when Anthropic and the Department of Defense entered a spat over the use of Anthropic's AI models in defense applications. This conflict led to Anthropic being dubbed a supply-chain risk by the administration.

• April 7: Anthropic introduced Mythos, a model adept at detecting and exploiting zero-day software vulnerabilities.
• April 30: The White House opposed expanding Mythos access, citing concerns about compute constraints.
• June 2: Anthropic announced that Project Glasswing partners had found over 10,000 serious software flaws and expanded the program to include 150 new organizations in 15 countries.
• June 9: Anthropic released Fable 5, a 'Mythos-class' model with stricter guardrails, claiming safeguards were sufficient for general release.
• June 10: Anthropic released frameworks calling for government regulation and legal authority to block risky models.
• June 11: Amazon's Andy Jassy informed the government that researchers found a way to get Fable 5 to provide information for cyberattacks.
• June 12: Senior White House staff met to discuss the situation.
• Thursday: Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent regarding jailbreak concerns.
• Friday Night: The government imposed stringent export controls, leading Anthropic to take the models offline entirely.

This timeline highlights the speed at which national security concerns can escalate into full-scale operational shutdowns. The conflict intensified after the administration learned that Mythos was made available to an entity in a foreign country with direct ties to the Chinese Communist Party, a factor that distinguished this incident from previous regulatory hurdles.

Why the White House Opposed Fable 5 Access

The mechanism behind the opposition to Fable 5 access involves a complex interplay of national security assessments and export control laws. The White House's stance was not merely a preference but a directive rooted in specific intelligence regarding the model's capabilities and the entities accessing it.

The process began with the administration's assessment of the compute constraints and the potential for these models to be used for cyberattacks. On April 30, the White House opposed expanding Mythos access, setting a precedent for caution. However, the situation escalated after the release of Fable 5 on June 9. Despite Anthropic's claims that their safeguards were "stricter than would be ideal" and sufficient for general release, the administration remained skeptical.

The core of the opposition lies in the specific vulnerabilities discovered. Amazon's researchers found a way to bypass the restrictions, allowing the model to "provide [it] with information that could be used to aid cyberattacks." This was supported by at least five other companies, including JPMorganChase, which participated in Project Glasswing. The administration's thinking, as noted by sources familiar with the situation, was that Anthropic had "come to every fork in the road and took the wrong fork" regarding safety protocols.

Furthermore, the administration threatened Anthropic with export controls after learning that Mythos was made available to an entity in a foreign country with direct ties to the Chinese Communist Party. This specific intelligence triggered the enforcement of existing export control frameworks. The administration viewed the distribution of such a powerful tool to a hostile actor as a critical failure in supply chain management. Consequently, the directive to suspend access by any foreign national was issued, effectively nationalizing the safety protocols by removing the models from the global market entirely.

Project Glasswing: Finding Flaws in Critical Software

Project Glasswing represents a critical initiative to harden software against AI-driven attacks. Launched to address the vulnerabilities exposed by the Mythos model, the project aims to detect and exploit zero-day software vulnerabilities before they can be weaponized.

How Project Glasswing Works The program utilizes AI models to scan for novel cybersecurity flaws. By leveraging the unique capabilities of models like Mythos and Fable 5, the project identifies weaknesses that traditional testing methods might miss. The goal is to reduce cybersecurity and biology-related risks by proactively finding flaws in critical infrastructure.

Practical Examples and Findings The efficacy of Project Glasswing was demonstrated through its rapid discovery of over 10,000 serious software flaws. These findings were expanded upon when the program grew to include 150 new organizations across 15 countries. The involvement of major financial institutions like JPMorganChase highlights the cross-sector nature of the threat.

Use Cases for Hardening Software

• Zero-Day Detection: Using AI to identify software vulnerabilities before they are exploited by malicious actors.
• Supply Chain Security: Scanning third-party libraries and dependencies for hidden flaws that could compromise entire networks.
• Biology Risk Reduction: Identifying potential biosecurity threats embedded in code or data processing pipelines.

Despite these efforts, the project faced significant challenges. Sources indicate that while the tools were effective at finding flaws, the communication gaps with the administration hindered the broader adoption of these safety measures. The administration argued that Anthropic failed to appreciate the ideological differences regarding how fast these tools should be deployed versus how safe they needed to be.

The Conflict: Communication Gaps and Ideological Differences

The Anthropic Fable shutdown cannot be understood without examining the breakdown in dialogue between the technology company and the government. A source familiar with the administration's thinking noted that Anthropic has not done a great job trying to "speak to the administration." This failure to communicate effectively became the central pillar of the conflict.

Ideological Differences The conflict was driven by differing views on AI safety and deployment. Anthropic prioritized safety and claimed their safeguards were sufficient for a general release. However, the administration held a stricter view, believing that the risks outweighed the benefits of a broad release. An administration official stated that Anthropic "failed to honor a recent cyber executive order," suggesting a fundamental disagreement on compliance and risk management.

Communication Breakdown Sources indicated that Anthropic did not appreciate the administration's ideological stance. The company chose to distribute the model knowing a jailbreak could happen, a decision the administration viewed as reckless. The phrase "They screwed us" was reportedly used in internal discussions regarding the lack of cooperation. This lack of cooperation extended to the refusal to fully comply with the spirit of the executive order until the government explicitly threatened export controls.

Comparison of Approaches

• Anthropic's Approach: Focus on releasing models with "safeguards that we (and, to our knowledge, all other AI developers) have yet to develop," even if they are "stricter than would be ideal."
• Administration's Approach: Prioritize national security and prevent access by foreign nationals, regardless of the technical claims made by the developers.

The administration's threat of export controls followed learning that Mythos was made available to an entity in a foreign country with direct ties to the Chinese Communist Party. This specific incident highlighted the gap between private sector risk assessment and government intelligence. While Anthropic revoked Mythos access for a global telecom company without the threat of export controls, the situation changed drastically when the threat level was perceived to be higher, leading to the Friday night shutdown.

FAQs on AI Export Controls and Jailbreak Risks

What triggered the immediate shutdown of Fable 5 and Mythos 5? The shutdown was triggered by a letter from the U.S. Department of Commerce directing Anthropic to suspend all access to the models by any foreign national. This followed the discovery that a jailbreak existed, allowing the models to provide information for cyberattacks, and the revelation that the models were accessible to entities with ties to the Chinese Communist Party.

Are there legal frameworks to block risky AI models? Yes, Anthropic released frameworks calling for government regulation and legal authority to block or deter the deployment of risky models. However, the current situation demonstrates that existing export controls and executive orders are being used to enforce these blocks when voluntary compliance is deemed insufficient.

What are the risks of jailbreaking AI models? Jailbreaking allows models to bypass safety guardrails, potentially providing information that could be used to aid cyberattacks. In the case of Fable 5, researchers found a way to get the model to provide such information. The risk is amplified when these models are accessible to foreign nationals or entities in hostile jurisdictions.

Frequently Asked Questions

Who is responsible for the Fable 5 shutdown?

The shutdown was a joint outcome of the U.S. Department of Commerce's directive and the subsequent enforcement of export controls. While Anthropic executed the technical shutdown, the decision was driven by government orders regarding national security and the prevention of access by foreign nationals.

How many flaws were found in Project Glasswing?

Project Glasswing partners found over 10,000 serious software flaws. The program expanded to include 150 new organizations in 15 countries to help harden critical software against these vulnerabilities.

What is the status of the Mythos model?

Anthropic has shut down all access to both Fable 5 and Mythos 5. Access was revoked for a global telecom company, and foreign national access was suspended entirely following the government's stringent export controls imposed on Friday night.

Sources

• The Anthropic Fable Mess — Hacker News
• Anthropic shuts down Mythos 5, Fable 5 due to government order - American Banker — Google News
• "They screwed us": Personality clashes sent Anthropic's models offline — Hacker News