OpenAI launches Astra AI model with a focus on cyber tasks

In short: OpenAI released a new AI model called Astra and is rolling it out first to cybersecurity customers, with wider access coming soon.

What happened

OpenAI launched Astra on Thursday. The company says it is its most capable AI model so far, and that it is better at using computers and web browsers to complete tasks.

Astra is available first to customers using Daybreak, OpenAI’s cybersecurity program. OpenAI says it will also roll out over the next week to its paid plans, including Pro, Plus, Enterprise, and Business, and it will be offered through its API (a way for other apps to plug into OpenAI’s tools).

OpenAI highlighted Astra’s security and coding skills. The company said it tested Astra on security “benchmarks,” which are standardized tests used to compare models, and claimed it can help find “zero-day” exploits (new security holes that are not publicly known yet), so defenders can fix them.

Why it matters

A model that is better at finding security holes could help protect software, but it could also raise concerns about misuse if the same skills are used for attacks.

Astra is also controversial because of a technique the article calls “opaque recurrence,” which can make it harder to see the model’s “chain of thought.” Chain of thought is the step by step reasoning an AI used, like showing your work in a math problem. OpenAI said monitoring how models reason is important, but it may get harder as models become more capable.

Source: TechCrunch AI