The trick: Zero Underneath
OpenAI hit the brakes on one cyber model and sold another one three days later.
Astra might be too dangerous to release. GPT-5.6-Cyber, trained to refuse less, is for sale now.
OpenAI said on August 7 it cannot rule out that its unreleased Astra model has 'critical' cyber capabilities, so it is slowing parts of the work.
Before you read on. Your call?
TRUE, BUT
0 scores published
the post publishes the threshold definition and exactly zero measurements. No scores, no success rates, no trial counts. An analyst quoted by CSO Online calls it a precautionary trigger, not a finished finding. Media turned that into 'OpenAI pauses Astra'. OpenAI paused only internal activities that did not yet meet new security controls.
The twist
three days later the same company expanded Daybreak and launched GPT-5.6-Cyber, trained to refuse less on 'higher-risk, dual-use cyber tasks'. It completes 95.0% of exploit-development requests, versus 1.5% for the base model. When OpenAI wants to publish numbers, it publishes numbers.
There’s more to this story.
Membership opens the full investigation, the strongest counterargument and what to do with what you’ve learned.
Start your free month →First membership: 30 days free, then A$89 a year. One introductory trial per customer. Card required; renews annually until cancelled. Cancel before the trial ends to avoid the first charge. Already a member? Sign in
Couldn't check your access. That's on us.
The trick has a name
We call it Zero Underneath: the headline number has nothing behind it. You'll see it again. Learn to spot it →
Receipts
- Supports openai.com:
we cannot rule out critical cyber capabilities under our Preparedness Framework
- Context openai.com:
pausing internal activities involving Astra that do not yet meet these strengthened security control requirements
- Refutes csoonline.com:
That is a precautionary trigger rather than a finished finding
- Context techcrunch.com:
critics have also pointed out that they function as marketing opportunities for the AI labs
- Supports helpnetsecurity.com:
cannot rule out the model reaching the critical capability level for cybersecurity
- Context openai.com:
completes 95.0% of these requests, compared with just 1.5%
Open the Receipts Pack → What each source proves, every figure traced, and what would change our verdict.