The US Government Tried to Restrict an AI Model. Cybersecurity Veterans Are Pushing Back
A week later, 76 veterans from the cybersecurity industry published an open letter demanding the government reverse the order. The signatories read like a who’s who of internet defensive infrastructure: Alex Stamos, former head of security at Facebook. Casey Ellis, founder of Bugcrowd. Jon Callas, a well-known cryptographer and former Apple security architect. Paul Vixie, a widely cited computer scientist. Dino Dai Zovi, previously head of application security at Block. Katie Moussouris, founder of Luta Security. Rachel Tobac, CEO of SocialProof Security. Their argument was direct: the government had handed the best AI security capabilities to adversaries at the exact moment those adversaries are building up their own offensive capabilities.
The trigger for the order appears to have been an internal report by Amazon researchers claiming they had found a way to bypass Fable safety restrictions to unlock Mythos-level capabilities. Anthropic confirmed to TechCrunch that the government order was likely based on this report. But Moussouris, who reviewed the report on her blog, argued that what the Amazon researchers described was not actually a jailbreak. They had asked Fable to fix open source code containing known vulnerabilities and to repair intentionally planted bugs. Fable accepted those requests. Moussouris concluded that this was not a safety bypass. This was the model’s core intended function: helping defenders find and fix vulnerabilities and test patches. Regulators had misidentified what the model was actually doing.
The open letter made another point that was more embarrassing for Anthropic: the technique described in the Amazon report could be replicated on essentially any other model, including OpenAI GPT-5.5, Anthropic’s own Claude Opus 4.8 and Sonnet, and even China Kimi 2.7. Moussouris told TechCrunch that Fable has restrictions that other models lack, so those other models do not need to bypass anything to perform the same security research tasks directly. Fable restrictions were actually stronger, not weaker, which made the export control rationale seem even thinner.
The letter closed with a policy ask: that AI regulation be developed through democratic legislative processes with transparent scientific input from industry and academic experts, rather than through ad hoc executive orders with unexplained justifications. Fable launched in April as a publicly available version following the Mythos preview. It had strict built-in guardrails preventing use in biology, chemistry, and cybersecurity domains, and preventing the model from being distilled into a replica of Mythos. But in practice, those restrictions were so broad that many legitimate cybersecurity researchers found their normal security-related queries getting rejected entirely. A model designed to be safer ended up being less useful for the exact researchers who most needed it.
Fable’s strict restrictions may have backfired in another way. By blocking so many legitimate security research queries, Anthropic may have inadvertently pushed researchers toward models with fewer safety features, which would actually be less safe for the tasks they were trying to perform. A model that says no to everything looks secure on a spec sheet. A model that approves security research requests and refuses malicious ones is harder to build, but considerably more useful in the real world. The export control debate may be a distraction from the harder question: how to build AI systems that can reliably distinguish between those two use cases, without blocking the former to prevent the latter.
