There you have it! OpenAI just labeled its own model "Critical" for cyber risk — the highest leve…
September 1, 2026 · 0 likes · 0 comments
Cybersecurity AI
There you have it! OpenAI just labeled its own model "Critical" for cyber risk — the highest level on its own scale — and plans to ship it anyway.
The model is called Astra. Per OpenAI's own blog post, it can find unknown flaws in hardened, real-world systems and write working exploits with nobody walking it through the steps. Perfect score on ExploitBench. In an internal test it discovered and exploited two zero-days on its own.
Read who's grading that. OpenAI built it. OpenAI decided it was safe. OpenAI is the only judge in the room. No named third-party audit. No federal standard to measure against — Trump's June executive order called for a voluntary review framework by August 1, and the White House still hasn't published it.
So the vetting is: trust us.
Now look at the same day. Anthropic shipped Fable 5.1 for coding and knowledge work, and put the dangerous cyber and bio capability in a separate model, Mythos 5.1, walled behind restricted-access programs and heightened safeguards. Same week, opposite instinct: gate the sharp end, don't hand it to everyone and hope.
Here's the rule nobody at these labs wants to hear: the entity that builds the weapon does not get to be the only one who certifies it's safe. That's not oversight. That's a press release.
A model that writes its own zero-days is a national security event, not a product launch.
Who audits the auditor before this ships?
Source: UnbiasedHeadlines.com
The model is called Astra. Per OpenAI's own blog post, it can find unknown flaws in hardened, real-world systems and write working exploits with nobody walking it through the steps. Perfect score on ExploitBench. In an internal test it discovered and exploited two zero-days on its own.
Read who's grading that. OpenAI built it. OpenAI decided it was safe. OpenAI is the only judge in the room. No named third-party audit. No federal standard to measure against — Trump's June executive order called for a voluntary review framework by August 1, and the White House still hasn't published it.
So the vetting is: trust us.
Now look at the same day. Anthropic shipped Fable 5.1 for coding and knowledge work, and put the dangerous cyber and bio capability in a separate model, Mythos 5.1, walled behind restricted-access programs and heightened safeguards. Same week, opposite instinct: gate the sharp end, don't hand it to everyone and hope.
Here's the rule nobody at these labs wants to hear: the entity that builds the weapon does not get to be the only one who certifies it's safe. That's not oversight. That's a press release.
A model that writes its own zero-days is a national security event, not a product launch.
Who audits the auditor before this ships?
Source: UnbiasedHeadlines.com