Topline
OpenAI will “soon” launch its most advanced cybersecurity AI model yet, according to a Tuesday statement, setting up the release of a direct competitor to Anthropic’s Mythos model, which rung alarms across financial world and government in April due to its high-powered autonomous cybersecurity features.
OpenAI did not give a date for Astra’s release.
Photo by Samuel Boivin/NurPhoto via Getty Images
Key Facts
OpenAI plans to launch Astra “soon,” though it noted the model’s “most advanced cybersecurity capabilities will be more limited,” with a select group of testers exclusively having access to them.
The model requires “stronger safeguards during development and before release,” OpenAI noted, saying Astra is now capable of finding unknown cybersecurity flaws and creating ways to exploit them autonomously “across many well-protected systems.”
The Astra announcement comes months after Anthropic launched a preview of Mythos, which was limited to vetted partners like Apple and Nvidia and created concern among observers as they realized its potential use for hacking financial services, operating systems and browsers.
OpenAI delayed parts of Astra’s development to bolster and test protections against cyber misuse, including training the model to reliably reject harmful requests.
Like Mythos, Astra can create zero-day exploits—security flaws unknown by developers—against real-world systems without human intervention, Tuesday’s statement said.
What To Watch For
It is unclear when exactly Astra will be launched. OpenAI says at launch, it expects Astra’s safeguards “to create more friction than we ultimately intend in order to protect against potential misuse,” adding Astra’s advanced cybersecurity features will first be made available to a small group of alpha testers.
Key Background
Though Astra will be a competitor to Mythos, the model is still months behind Mythos’ releases. The preview for the Anthropic model launched in April and had its first public version released in June. Anthropic said Mythos would incorporate new features to detect misuse and that the company was working with the federal government to “steadily expand access” to the more advanced model of Mythos. Both Mythos and Astra can outperform human cybersecurity workers in efficiency, though OpenAI and Anthropic have acknowledged humans are needed to correct mistakes and confirm real-world context the models can potentially have trouble registering.
Further Reading
Anthropic Releases First Public Version Of Claude Mythos—With Major Safeguards (Forbes)
Be the first to comment