TOBIAS SCHOTTSTÄDT
FULL-STACK DEVELOPER
FROM KASSEL

The recent global suspension of Anthropic’s new AI model, Fable, is reportedly attributed to the influence of another major tech company: Amazon CEO Andrew Jassy. According to these reports, he is said to have informed US Treasury Secretary Scott Bessent of a suspected ‘jailbreak’. This allegedly led to the US government issuing an export directive, which ultimately forced Anthropic to block the AI model.

Various media outlets are reporting the same story: Andrew Jassy is said to have initially instructed Amazon’s cybersecurity department to test the new AI model for potential ‘jailbreaks’ – methods used to trick AIs into processing sensitive prompts that would normally be rejected. The experts are said to have submitted a report revealing working exploits for security vulnerabilities in four common programmes and informed Jassy. He is also said to have shared this information with US government officials.

These reports were followed by an ultimatum from the US government to Anthropic: the jailbreak issues must be resolved within 24 hours, otherwise access to Fable would be blocked for non-US citizens. David Sacks, a tech investor and co-chair of the US President’s Advisory Council on Science and Technology, reported this on X (Twitter). According to his account, Anthropic CEO Dario Amodei is said to have refused to patch the security vulnerabilities resulting from Amazon. Although Anthropic had previously denied that this constituted a full-scale jailbreak, according to Sacks’ account, the strict ban was imposed due to Amodei’s refusal.

According to experts, the US government’s approach is illogical. In theory, any AI model can be compromised by a jailbreak. Furthermore, Anthropic itself had attracted considerable attention because Fable (a variant of the Mythos model) was reportedly able to identify previously unknown security vulnerabilities in almost any software in a very short time and provide exploit code for them – a capability that adversaries deemed too dangerous for general release. The current move could therefore also be seen as an indirect criticism within an existing conflict between Anthropic and the US government.

Source: www.heise.de