OpenAI’s Astra mannequin has AI researchers spooked. Right here’s why
Arrived simply days after the launch of Anthropic’s Claude Fable 5.1 and Mythos 5.1, Astra marks a “soar” in AI capabilities, OpenAI president Greg Brockman mentioned, boasting that the brand new mannequin “can actually do something a human can do with a pc.”
Astra can also be OpenAI’s first mannequin to succeed in the “essential” threshold of the corporate’s “preparedness framework” as a result of its excessive cybersecurity expertise, that means it might perform “end-to-end” assaults on “hardened targets” by itself, amongst different capabilities.
OpenAI beforehand paused work on Astra to bolster its safeguards earlier than asserting earlier this week that the mannequin is “persistently extra more likely to respect express security restrictions and warnings” than GPT-5.6 Sol, the OpenAI mannequin concerned within the now notorious Hugging Face assault.
Regardless of OpenAI’s assurances, AI consultants stay fearful about Astra. The brand new mannequin is claimed to make use of a reasoning approach recognized variously as “recurrent depth” or “opaque recurrence,” which (as TechCrunch describes) makes its “chain of thought” a lot tougher to learn.
Preserving tabs on a frontier AI mannequin’s pondering is, clearly, a giant deal in the case of stopping the sorts of rogue AI hacks we’ve been listening to about over the previous a number of weeks, and the potential of dropping that form of surveillance has spooked high AI researchers.
“If that is true, OpenAI appears to be violating one of many few redlines that exist within the AI neighborhood,” wrote Steven Adler, a former OpenAI security lead, on X.
