Preparing our next security sprint to harden OpenAI
Nov 2, 2026, 10:00 AM · OpenAI
OpenAI is turning its own models and an outside firm loose on its defenses again, and the approach is sound, though the post is long on intent and short on measurable results.
Why it matters
OpenAI says it's planning another internal security sprint, this time bringing in the security firm Cathedral, whose researchers will use OpenAI's models alongside its own teams to find weaknesses. It's paired with a bigger bug bounty, with critical rewards more than quadrupled.
The company says its last sprint put 25% of its product engineers on finding and fixing weaknesses, using Codex to investigate, reproduce bugs in isolation and prepare patches. It also points to $1 billion in subsidized access over six months for frontline defenders of essential services.
The argument underneath all this is that attackers now move at AI speed, so defenders have to as well.
From the desk
We think this is the right instinct. If frontier models can find bugs faster, the people building them should be the first to aim that at their own systems. Bringing in outsiders with different methods is also smart, since internal teams tend to miss the same things over and over.
Putting a quarter of product engineering on security in the last sprint is a serious commitment, and we credit it. Few companies pull that many builders off shipping.
What's missing is the scorecard. The post says the last sprint showed what the approach can deliver, but it doesn't say how many issues were found, how severe they were, or how many were fixed. Without that, it reads more like a statement of posture than a progress report. OpenAI mentions sharing architecture and lessons elsewhere; we'd like numbers next time.
There's also a commercial thread. The same models doing the defending are sold to Daybreak customers, and the post doubles as a pitch. That's not disqualifying, but readers should hold it in mind. The case for AI-driven defense is strong. It's also a market.
If this scales, the likely outcome is an arms race where defenders who can afford frontier tools pull ahead and everyone else falls further behind. The subsidized-access commitment is meant to address that. Whether it reaches the under-resourced operators who need it most is what I'm watching.
Context
This announcement comes alongside OpenAI's separate post expanding its Security Bug Bounty. The company frames both under what it calls keeping the defender's window open, citing recent incidents across the industry as evidence that threats now emerge quickly.
Who feels it
- Security researchers
- More paid work and outside partnerships are on the table, both through Cathedral-style engagements and the bounty.
- OpenAI customers
- Continuous testing should mean fewer exploitable gaps in the products they rely on, though there's no public metric yet to confirm it.
- Critical infrastructure defenders
- The subsidized access could bring frontier tools to teams that couldn't otherwise afford them.
- Other AI labs
- A public bar is forming for using your own models against your own stack.
What to watch
- Whether OpenAI publishes concrete results from the last sprint and this one.
- Who actually receives the subsidized defender access, and how it's allocated.
- What Cathedral's researchers find, and whether any of it is disclosed.
- Whether other labs announce comparable red-team sprints with outside firms.
Companies: OpenAI