Feds freaked over Fable 5 after 'fix this code', not jailbreak, say researchers (theregister.com)

Article summary

The US government blocked access to Anthropic's Fable 5 and Mythos 5 models, citing national security concerns, after a research paper claimed the models could be used to bypass security guardrails. However, a researcher who read the paper claims the 'jailbreak' was actually just a simple prompt to 'fix this code'. The researcher argues that this capability is essential for defensive cybersecurity and that removing it would make the models less useful for finding and fixing bugs. The incident has sparked debate about the balance between security and the potential benefits of AI models like Fable 5.

Main themes

• AI security
• National security
• Cybersecurity
• AI regulation
• Export controls

What commenters say

• The US government's decision to block access to Fable 5 and Mythos 5 was likely an excuse to restrict access to the models rather than a genuine security concern.
• The ability to ask AI models to find and fix bugs is essential for defensive cybersecurity and should not be restricted.
• It is impossible to create AI models that can discern intent and prevent malicious use, and therefore, restrictions on their use are futile.
• The focus on restricting AI models is misguided, and instead, efforts should be made to improve cybersecurity practices and protect against potential threats.
• The use of AI models like Fable 5 and Mythos 5 can actually improve security by identifying vulnerabilities and allowing for faster patching.
• The restriction on access to Fable 5 and Mythos 5 will ultimately harm defenders more than attackers, as it will limit their ability to use the models for cybersecurity purposes.
• The development of AI models like Fable 5 and Mythos 5 is a cat-and-mouse game, where the models will always be one step behind potential attackers, and restrictions will only slow down progress.
• The idea of using ID verification to restrict access to AI models is flawed, as it is easy for malicious actors to obtain fake identities or exploit other vulnerabilities.