The U.S. Department of Commerce has imposed export controls on Anthropic’s advanced AI model, Fable 5, a move that has sparked debate within the cybersecurity and artificial intelligence communities. This decision, enacted last Friday, aims to prevent potential misuse of powerful AI capabilities by foreign adversaries.
Anthropic, the AI company behind Fable 5, had previously implemented safeguards to mitigate risks, including not releasing its most capable model, Mythos, publicly and developing specific guardrails for Fable 5 to limit responses on sensitive topics like cybersecurity and biological warfare. Despite these measures, the administration expressed concern following reports from Amazon and other researchers claiming they had successfully bypassed Fable 5’s protections shortly after its release.
Commerce Department Restricts Fable 5 AI Model
The Department of Commerce’s action led to Anthropic temporarily disabling Fable 5 for all users while the company engaged with the White House. The administration’s rationale, according to reports, was that if U.S. researchers could circumvent the model’s safeguards, then foreign entities could also achieve this, posing a national security risk.
However, this decision has met with significant criticism from cybersecurity and AI experts. Many argue that the reported bypasses do not demonstrate the ability to circumvent Fable 5’s core safety features or unlock dangerous new capabilities that would warrant such severe export controls.
Expert Disagreement on AI Model Safeguards
Katie Moussouris, a prominent cybersecurity expert, stated on Monday that Anthropic provided her with third-party research detailing methods used to interact with Fable 5’s guardrails. According to Moussouris, researchers used a multi-step process to prompt Fable 5 to analyze vulnerable open-source code for security issues. While Fable 5 initially declined the request, the process eventually led the model to generate scripts that could test patches for the identified vulnerabilities.
Moussouris argues that these demonstrated capabilities are fundamental to Fable 5’s value for cybersecurity defense, not a bypass of its intended safeguards. She explained that defenders need AI tools to identify bugs, explain their significance, and create tests to confirm fixes, which are standard practices in the cybersecurity field.
“Defenders need to be able to ask AI to fix the bugs in a file, explain why the fix matters, and write tests that confirm the patch works,” Moussouris stated. “That is not a guardrail bypass. It is the most valuable thing an AI model can do for defensive security: executing the find, fix, and test loop defenders run every day.”
Drawing on her experience with the Wassenaar Agreement, a multilateral export control regime, Moussouris described the restrictions on Fable 5 as “heavy-handed” and “misguided” based on the research she has reviewed.
Anthropic’s Testing and Wider Community Concerns
Anthropic itself reported that extensive red team testing, totaling 1,000 hours, by both internal and external experts did not reveal universal jailbreaks that would remove Fable 5’s guardrails or grant access to its most powerful Mythos capabilities for sensitive applications.
Moussouris is among dozens of cybersecurity experts who signed an open letter urging the Trump administration to reconsider its decision, using the slogan “Free Fable.” The letter highlights that while Mythos-class models are proficient at identifying software vulnerabilities, their performance is comparable to other advanced models used daily for cybersecurity defense. For instance, OpenAI’s Daybreak model, which possesses similar capabilities, was not subject to Commerce Department restrictions.
The researchers also pointed out that Fable 5’s guardrails were perceived as overly cautious compared to other frontier models, humorously hindering basic defensive cybersecurity tasks for some users upon its release. The open letter further questions whether the reported bypasses constitute true offensive capabilities, noting that similar functionalities can be replicated by other commercially available and open-source models, including GPT 5.5, Claude Opus, Claude Sonnet, and Chinese models like Kimi 2.7.
The letter emphasizes that AI has been capable of finding bugs and generating exploits for some time, challenging the notion that Fable 5 offers a unique “uplift” in these areas to justify the unprecedented export control action.
Broader Context of AI Regulation
The White House’s decision on the Fable 5 AI model occurs against a backdrop of increasing public demand for stronger government intervention in AI development and deployment. Recent polls indicate broad bipartisan support for AI regulations.
A Johns Hopkins University poll from May found that a significant majority of Americans favor regulations such as bans on AI-generated images and video, mandatory labels on AI content, and disclosure laws for AI interactions. Additionally, a global survey highlighted public concerns centered on AI’s potential to spread misinformation, generate deepfakes, facilitate cybercrime, and aid in the creation of new weapons.
The next steps are uncertain, with Anthropic continuing to engage with the White House. The duration of the export controls and the possibility of further regulatory actions on advanced AI models remain areas to watch in the evolving landscape of artificial intelligence governance.