Anthropic has pulled its latest language model from service after a US government order citing possible jailbreaking risks. The move stops widespread use of Fable 5 and Mythos 5, undermining plans to deploy the system in climate modelling and environmental monitoring. While regulators cite security, critics argue the halt slows the potential of advanced AI to accelerate climate solutions and reduce emissions. The firm said it had designed safeguards ahead of release, yet the black‑listing underscores how re‑security checks can postpone tech that could help track changes in global systems. A judge ruled the Pentagon’s restriction cannot be enforced, but the legal battle shows the tension between national security and innovation.
In a statement on its website Anthropic said the order required a suspension of foreign use of Fable 5, a program it called "too powerful". The halting of the tool means customers cannot access the model until the security assessment is resolved, which may take months. That pause also delays the use of the AI for sensitive climate data analysis, making it harder for governments to monitor deforestation patterns or model carbon flows at scale.
Security experts note that the US government believes it has discovered a bypass technique that lets users circumvent the model’s safety filters. Anthropic reviewed a demonstration of the attempted jailbreak and found it exploited minor vulnerabilities which other publicly available models could find without bypassing safety measures. “These vulnerabilities all appear relatively simple,” the firm said. “We have found that other models discover them as well.”
In line with its earlier pre‑release test, Anthropic offered the AI to a select group of organisations to present the model’s intelligence while safeguarding data. The company said it was confident its safeguards would prevent cyber hacking, yet the new order suggests those safeguards fell short in the context of national security.
Environmental policy experts point to the impeding effects of such restrictions on climate research. Gina Neff, Professor of Responsible AI at Queen Mary University London, warned that limiting the model could delay the development and safe testing of AI systems for environmental purposes and hamper collaboration across governments. “We’re in uncharted territory at this point,” she said. “People within the AI industry have been warning that these tools are getting better very rapidly and that we have to be able to build up capabilities to keep our companies safe from cyber attacks.”
The conflict also intersects with broader regulatory scrutiny. The US president’s administration recently labelled Anthropic’s models a supply‑chain risk, the first time that classification was applied to an American technology company. The designation would normally exclude the tool from government use, but a court has so far blocked enforcement of the Pentagon’s directive, allowing agencies still the option to use the AI while the lawsuit continues.
While the legal challenge plays out, the environmental community watches closely. The potential for advanced language models to analyze vast environmental datasets – from satellite imagery to emissions records – is immense. The current pause demonstrates how regulatory stances can affect the delivery of AI innovations that might help nations track climate change or devise mitigation strategies. Future policy will need to balance security concerns with the urgent need for tech tools to manage and reverse environmental degradation.
