OpenAI has unveiled GPT-6 Astra, a new version of the AI model behind ChatGPT 

The company says this is a major leap in capability, including in coding, “computer use”, science, and cybersecurity. OpenAI’s president told media publications that Astra’s launch could be viewed as a possible step toward artificial general intelligence (AGI). 

The All Ireland Science Media Centre asked local experts to comment.

Maribeth Rauh, Researcher, AI Accountability Lab, Trinity College Dublin, comments: 

Do claims about a possible step toward artificial general intelligence (AGI) stand up to scrutiny? 

“‘AGI’ continues to be an ill-defined, movable goalpost which only serves marketing purposes. Even if we were to put stock in such a concept, the ability to do software engineering tasks and use a browser is far narrower than what could constitute any notion of “general” or human intelligence. Rather, these tasks are those which AI developers are most familiar with and place most value on. Moreover, these tasks are among the most amendable to automation and verification via objective, standardised measures, unlike many other facets of engineering and scientific work, and indeed most human activity.”

Do you have any cybersecurity, privacy, or broader safety concerns about the release of GPT-6 Astra?

“OpenAI refers to this model as their “most aligned,” substantiated by measures of their ability to contain the model when it attempts programming tasks. What happened to alignment with some notion of societal benefit, or even minimization of the multitude of other harms this technology fuels: from psychological harm, fraud, at-scale misinformation, and AI slop, to impacts on the environment and economic or educational systems? It’s a striking example of just how much any notion of real benefit has been eroded.

“Given the lack of regulation and enforcement, the extent of the privacy and cybersecurity risk relies heavily on OpenAI’s actions. They need to commit to proceeding slowly enough to implement existing best practices and invest time and effort to develop new guardrails.  Their track record, and the track record of the tech industry writ large, doesn’t give me confidence they will do that; I would advise against giving GPT-6 Astra access to sensitive personal information or systems, no matter how alluring the supposed benefits.

“The fact that GPT-6 Astra scored a perfect 0% on their new “ExploitGym” evaluation only suggests that they’ve trained the model to ace that specific test. It isn’t necessarily a meaningful measure of their ability to secure the model more generally. It also underscores how mitigating the risks of these models is a game of whack-a-mole. There are likely many other ways the model could cause cybersecurity issues which this benchmark does not measure. Stay tuned for the next benchmark for a different flavour of cybersecurity risk.”

Declaration of interest: “I was formerly employed by Google DeepMind as a Research Engineer working on ethics and safety. I worked there from 2018 – 2024.”