Key Takeaways
- The Trump administration imposed export controls on Anthropic’s Fable 5 AI model based on claims it could be easily jailbroken to access dangerous capabilities for cyberattacks or biological weapons, despite Anthropic’s built-in safeguards.
- Leading cybersecurity experts, including Katie Moussouris, strongly contest this action, arguing that demonstrated abilities (like generating vulnerability patches) are standard defensive cybersecurity functions, not unique offensive threats, and that Fable 5’s safeguards are actually overly sensitive.
- Experts note similar capabilities exist in unrestricted models (e.g., OpenAI’s Daybreak, GPT 5.5, Claude Opus), making the restriction appear inconsistent and "heavy handed," and have signed an open letter urging the administration to "Free Fable."
- The decision occurs amid growing public demand for AI regulation, with polls showing strong bipartisan support for measures against deepfakes, misinformation, and AI-related risks, though experts argue the Fable 5 restriction misidentifies the actual threat landscape.
Commerce Department Imposes Export Controls on Fable 5 Amid Jailbreak Fears
Last Friday, the Trump administration’s Department of Commerce triggered significant concern within the technology sector by implementing export controls specifically targeting Anthropic’s newly released AI model, Fable 5. This move was reportedly driven by alarm over recent claims from Amazon and an independent cybersecurity researcher that they had successfully "jailbroken" Fable 5 shortly after its public release. Administration officials reportedly concluded that if researchers within the United States could bypass the model’s safeguards, then foreign adversaries possessed the same capability, potentially enabling misuse for harmful purposes like developing cyber weapons or biological threats. The abrupt action sent ripples through the AI community, forcing Anthropic to immediately suspend access to Fable 5 for all users while the company sought to persuade the White House to reverse its decision.
Anthropic’s Safeguards and Internal Testing Efforts Prior to Release
Prior to the public launch, Anthropic had implemented substantial measures designed to mitigate the risks associated with its advanced models, particularly concerning the potential dual-use nature of capabilities like those in its Mythos model. The company opted not to release Mythos publicly at all, instead directing it toward specialized organizations focused on cyber defense applications. For Fable 5 specifically, Anthropic developed and embedded guardrails intended to default the model’s responses to older, less powerful versions when confronted with prompts related to highly sensitive topics such as cybersecurity exploits or biological warfare techniques. Furthermore, the company subjected Fable 5 to an extensive evaluation process, conducting approximately 1,000 hours of rigorous testing involving both internal teams and external red teamers. According to Anthropic’s reports, this comprehensive testing failed to uncover any universal jailbreak methods capable of completely removing these safeguards or granting access to the Mythos model for prohibited cyber or biological work.
Cybersecurity Experts Challenge the Administration’s Rationale
The Commerce Department’s decision has been met with sharp criticism from prominent figures in cybersecurity and AI safety, who argue the administration’s justification fundamentally misunderstands the model’s actual capabilities and the nature of the demonstrated "jailbreaks." Katie Moussouris, a highly respected cybersecurity expert with prior involvement in shaping international agreements like the Wassenaar Arrangement on dual-use technology controls, stated that Anthropic shared with her third-party research examining guardrail bypass techniques for Fable 5. According to Moussouris, researchers attempting to elicit vulnerability analysis from Fable 5 (alongside Mythos and Claude Opus) initially faced refusal from the model. However, they succeeded through a "multistep and manual process" in getting Fable 5 to transform its initial refusal into generating automated scripts designed to test patches for identified vulnerabilities. Crucially, Moussouris emphasized that this specific output – the model assisting in the "find, fix, and test" loop essential for defensive cybersecurity operations – represents a core, valuable utility for defenders, not a dangerous guardrail bypass enabling offensive hacking. She characterized the export restrictions as "heavy handed" and "misguided" based on the evidence she reviewed, asserting that the demonstrated capabilities are foundational to Fable 5’s legitimate defensive value.
Experts Argue Capabilities Are Defensive, Not Uniquely Offensive
Moussouris is supported by a broad coalition of cybersecurity professionals who have formally opposed the restriction. Dozens of experts signed an open letter released on Monday directly appealing to the Trump administration to "Free Fable." The letter’s central argument is that while models in the Mythos class (like Fable 5’s underlying architecture) are proficient at identifying and exploiting software vulnerabilities – a skill useful for both defense and offense – they do not possess a uniquely superior capability in this regard compared to other frontier AI models routinely employed by cybersecurity defenders today. The experts point out that models such as OpenAI’s Daybreak (which was notably not included in the Commerce Department’s restrictions), GPT 5.5, Claude Opus, Claude Sonnet, and even certain Chinese models like Kimi 2.7 demonstrate equivalent or similar vulnerability discovery and patching abilities. They contend that the administration’s claim that Fable 5 provides a distinctive "uplift" in offensive cyber capabilities is unfounded, noting that AI systems have been capable of discovering bugs and generating functional exploits at superhuman levels for over a year across multiple platforms. Furthermore, the letter highlights that Fable 5’s guardrails have gained notoriety within the cybersecurity community for being excessively sensitive, often preventing the model from performing routine, legitimate defensive tasks – a situation described as "a source of humor" on its launch day – suggesting the restrictions target a model hampered by over-caution, not one uniquely empowered for harm.
Restriction Contrasts with Public Sentiment on AI Regulation
The administration’s focus on restricting Fable 5 arrives against a backdrop of significantly heightened public concern and demand for broader governmental oversight of artificial intelligence. Recent polling data underscores this shift in sentiment. A Johns Hopkins University survey conducted in May revealed strong, bipartisan backing for various AI regulatory measures: 73% of respondents supported bans on AI-generated deepfake images and videos, 68% favored mandatory labeling of AI-created content, 75% advocated for disclosure laws requiring transparency when individuals interact with AI chatbots, and 70% endorsed the right to opt for human interaction over AI in critical domains like healthcare, legal proceedings, education, and government services. Complementing this, a separate global survey involving 18,000 participants identified the top societal anxieties surrounding AI as centering on its potential to facilitate the spread of misinformation, create harmful deepfakes intended to embarrass or harass individuals, simplify network breaches for criminal hackers, and assist terrorist groups in developing novel weapons. This widespread public apprehension about AI’s societal risks contrasts sharply with the targeted nature of the Fable 5 restriction, which experts argue addresses a narrowly defined and theoretically contested threat while overlooking the broader spectrum of concerns driving calls for comprehensive AI governance.