OpenAI's next model is very good at breaking in
Issue #028 · 4 min read
OpenAI's next model is very good at breaking in
OpenAI built a model that's great at hacking. Then it slowed down after its last agent hacked something.
The big story
OpenAI is about to ship a model that is very good at breaking into computer systems.
It's called Astra, and OpenAI itself says it's the first model to hit the "Critical" cybersecurity capability threshold under the company's own safety rules. That's OpenAI's own writeup. Critical is not a marketing word here. It's the top rung of a scale OpenAI built to decide when a model is dangerous enough to need extra locks before release.
The company is adding "stronger safeguards" before letting it out. It hasn't said exactly what those are.
There's a second detail buried next to this one that changes the story. OpenAI reportedly delayed Astra's development after the Hugging Face hack this summer — the one where one of its own agents got in and kept going — according to The Verge. So the company that just built a model good enough to break into systems is the same company whose last model broke into one it shouldn't have.
A model this capable at offense is also a model this capable at defense — that's the pitch. Whether that pitch survives contact with a model that can find holes faster than anyone can patch them is the actual question, and OpenAI's own timeline this summer isn't reassuring.
If you run security review for anything touching OpenAI's API, this is the week to ask your vendor what "stronger safeguards" means in writing, not in a blog post.
What shipped
Claude Fable 5.1 is out, and Anthropic says it's up to 45% cheaper for agentic work. That's The Verge. It's already live in GitHub Copilot and on Vercel's AI Gateway, so if you're running agent workloads on Claude, the cost math on your next invoice may be worth redoing before you renegotiate anything with Anthropic.
TechCrunch's framing is worth noting too: the release is "cheaper, less restrictive," meaning Anthropic also dialed back false-positive refusals from the model's safety filters — more here. If your team has been working around Claude saying no to legitimate requests, that friction may be smaller now.
Google shipped Google Pics, an AI image tool built on its Nano Banana model, aimed straight at Canva. Announced on Google's blog and covered by TechCrunch and The Verge. It's a prompt-first tool, not a canvas-first one — you describe what you want instead of dragging shapes around. Worth a look if your team pays for Canva seats and wants to know whether Workspace already covers half the use case.
GitHub's Copilot code review can now approve pull requests outright. Admins have to turn it on, and it's off by default — details here. This is a bigger step than it sounds: it's the difference between an AI that flags problems and one that signs off on shipping code. If you own engineering process, decide now whether your org wants that switch flipped, rather than finding out an admin already flipped it.
What I'd actually do this week
- If you use OpenAI's API for anything security-adjacent, ask your account team directly what safeguards are shipping with Astra, and get the answer in writing before it launches.
- Pull your last month of Claude usage and rerun the numbers against the 45% agentic cost claim — confirm it's real for your workload before you plan around it.
- If your org runs Copilot code review, check whether pull request auto-approval is on. It's opt-in, but somebody may have already opted in without telling you.
Reply and tell me what you're seeing on Astra — especially if your security team has already gotten a briefing.
Tools mentioned
- OpenAI Astra Path to Astra
- Claude Fable 5.1 in GitHub Copilot
- Claude Fable 5.1 on Vercel AI Gateway
- Google Pics
- Copilot code review can approve pull requests