OpenAI Unveils GPT‑6 “Astra” with Built‑In Cybersecurity Guardrails

OpenAI slipped a brand‑new AI engine into the hands of a handful of testers on Thursday, calling it GPT‑6 “Astra”. It isn’t available to the general public just yet – only firms that are part of the company’s Daybreak program, plus paid ChatGPT users who opt‑in for the extra‑secure version.

Greg Brockman, the firm’s president, used a briefing to note that Astra could eventually be seen as one of the first glimmers of artificial general intelligence – that elusive software that can outperform humans across a wide gamut of tasks. He was careful, though, to add that there’s still “a lot of work to be done”, and that whether it truly qualifies as AGI is something each reader will have to decide for themselves.

Why the fanfare? For starters, Astra is noticeably better at taking over a computer and doing work on a user’s behalf. Think longer, more involved tasks – drafting a report, analysing a dataset, even troubleshooting a piece of code – something earlier models could only nibble at.

But the rollout isn’t just about raw capability. After a series of unsettling incidents – notably an accidental hack of Hugging Face in July and a wave of AI‑powered attacks on water‑treatment facilities – OpenAI has doubled down on safety. Both the Daybreak‑only version and the paid‑ChatGPT version come with hardened guardrails that block the model from accessing its most powerful cybersecurity tricks.

The company says Astra hit what it calls a “critical cybersecurity threshold”, meaning it can, in theory, discover and even craft zero‑day exploits without human help. To keep that power in check, OpenAI paused some internal work on the model back in August, added extra safeguards, and now promises a version that can’t be misused for malicious hacking.

At the same conference, Brockman announced a new initiative to give utilities, local governments and financial institutions subsidised access to the Daybreak cybersecurity program. He also hinted at a multi‑state information‑sharing network aimed at spotting AI‑driven threats before they cause damage.

Sam Altman, the CEO, admitted that Astra’s launch took longer than the team hoped. “We wanted to make sure the guardrails were solid before we let more people play with it,” he told Bloomberg TV.

In short, OpenAI is trying to walk a tightrope: push the envelope on what AI can do, while stitching up the safety net that recent events have shown is too often left gaping. Whether Astra truly marks a shift toward AGI, or simply a better, safer assistant for businesses, will become clearer as more hands get to test it.