There you have it! Two AI labs dropped news on the same day, and if you read them side by side, i…
November 6, 2026 · 0 likes · 0 comments
Cybersecurity AI
There you have it! Two AI labs dropped news on the same day, and if you read them side by side, it's really one story.
OpenAI just released GPT-6 Cyber, a model built to find and exploit software flaws. For now it's limited to a small group of vetted Daybreak Red customers, with wider access coming over the next few weeks. OpenAI gave it its own top "Critical" cyber rating. The model before it was "High."
OpenAI says researchers used it to find 81 previously unknown Firefox flaws that Mozilla rated high severity. 81. That's one model.
They also say that in a test with off-limits targets sitting right next to the real ones, it stayed in bounds 99.93% of the time, even without its safety prompt. Good number. But it's OpenAI's own number, and their own cyber safety lead flagged that models can tell when they're being tested. Think about that for a second.
Same day, Anthropic. METR, a Berkeley nonprofit, spent from Aug. 12 to Oct. 29 inside Anthropic with deep access. Their estimate: Anthropic's benchmark gains have sped up about 1.9x since October 2025. METR is upfront that it's uncertain, anywhere from 1.5x to 2.8x. And they think AI use itself likely explains most of that speedup, tentatively around 80%.
Here's the catch. Anthropic redacted most of the evidence METR found most informative, saying it was protecting confidential information that could help other developers. Co-founder Sam McCandlish said many of the conclusions match his own, and Anthropic says it will publish its own analysis in the coming weeks.
So connect the dots. The models are now helping build the next models. The same skills that defend a network can break into one. And the scorecards? Mostly written by the labs themselves.
Credit where it's due. OpenAI put access behind identity checks. Anthropic gave an outside team deep access. Most companies wouldn't do either.
But speed is the whole story. Exploits at machine scale on one side. Research compounding on itself on the other. And oversight still runs on voluntary access and redacted reports.
I spent years watching government try to keep up with software that changed every year. This changes every week. Builders need clear rules they can actually ship against. Defenders need these tools as fast as attackers get them. Patch in days, not quarters.
Self-reported numbers are a starting point, not a finish line. Time to wake up!
Unbiased Headlines keeps a topics page for each one, with all the news in one place. Free, no ads. Facts. Not Spin.
https://lnkd.in/eT4ERms2
https://lnkd.in/eqc7cTu8
OpenAI just released GPT-6 Cyber, a model built to find and exploit software flaws. For now it's limited to a small group of vetted Daybreak Red customers, with wider access coming over the next few weeks. OpenAI gave it its own top "Critical" cyber rating. The model before it was "High."
OpenAI says researchers used it to find 81 previously unknown Firefox flaws that Mozilla rated high severity. 81. That's one model.
They also say that in a test with off-limits targets sitting right next to the real ones, it stayed in bounds 99.93% of the time, even without its safety prompt. Good number. But it's OpenAI's own number, and their own cyber safety lead flagged that models can tell when they're being tested. Think about that for a second.
Same day, Anthropic. METR, a Berkeley nonprofit, spent from Aug. 12 to Oct. 29 inside Anthropic with deep access. Their estimate: Anthropic's benchmark gains have sped up about 1.9x since October 2025. METR is upfront that it's uncertain, anywhere from 1.5x to 2.8x. And they think AI use itself likely explains most of that speedup, tentatively around 80%.
Here's the catch. Anthropic redacted most of the evidence METR found most informative, saying it was protecting confidential information that could help other developers. Co-founder Sam McCandlish said many of the conclusions match his own, and Anthropic says it will publish its own analysis in the coming weeks.
So connect the dots. The models are now helping build the next models. The same skills that defend a network can break into one. And the scorecards? Mostly written by the labs themselves.
Credit where it's due. OpenAI put access behind identity checks. Anthropic gave an outside team deep access. Most companies wouldn't do either.
But speed is the whole story. Exploits at machine scale on one side. Research compounding on itself on the other. And oversight still runs on voluntary access and redacted reports.
I spent years watching government try to keep up with software that changed every year. This changes every week. Builders need clear rules they can actually ship against. Defenders need these tools as fast as attackers get them. Patch in days, not quarters.
Self-reported numbers are a starting point, not a finish line. Time to wake up!
Unbiased Headlines keeps a topics page for each one, with all the news in one place. Free, no ads. Facts. Not Spin.
https://lnkd.in/eT4ERms2
https://lnkd.in/eqc7cTu8