GPT-6 Cyber is our latest model for vulnerability research and redteaming
Combining Astra’s innate capabilities with training designed and inspired by the work of our vulnerability research team including @ky1ebot@daveaitel@halvarflake@clintgibler and many others!
We found that GPT-6.2 was another step up in vulnerability research akin to the leap we saw from 5.6 -> 6
On unseen v8 bugs: an increase of 22%!
Along with this, our team worked hard to bring large improvements in secure code generation and secure and complete patch generation
Introducing GPT-6 Cyber! We trained it for advanced cybersecurity work with improved alignment. It has been powering our security work for quite a while and has been behind many of our recent security highlights: finding and helping patch high-severity vulnerabilities in Firefox, developing exploits and proposing patches for Linux kernel vulnerabilities, and strengthening OpenAI’s own infrastructure through red-teaming.
Today, we’re rolling it out to Daybreak Red to keep the cyber frontier in defenders’ hands.
openai.com/index/introducing…
We are continuing grow our vulnerability research team with new amazing talent such as @maddiestone@jeffreycrowell@_sy1vi3@jeffmcjunkin
I am grateful to be able to learn so much from everyone and to use what I learn to make our cyber models even better!
Great work by @ky1ebot
The most interesting part of this was seeing the model discover new previously unknown exploitation techniques
Since we ask it to exploit every single bug it is stuck with some very constrained limited bugs which require very creative thinking to pull off
We are now allocating a significant amount of our frontier compute towards discovering and patching vulnerabilities in open source codebases, fully pro bono.
In one case, we studied the Linux kernel, which is a particularly load-bearing piece of software at OpenAI.
Before AI, security researchers might discover ~20 local privilege escalation (LPE) vulnerabilities in Linux in a typical year.
In one week's time, our upcoming frontier cyber model was able to find ~1000 LPE vulnerabilities using about $1M of tokens. In the process, it was able to discover at least 10 novel classes of Linux exploitation techniques.
Firefox 159.0 just fixed 41 sechigh vulnerabilities found by me and @ky1ebot
Most of these leads to code exec in the renderer!
Along side a heroic effort from the Mozilla security team to secure Firefox with agents!
One of many projects we are working to harden, more to come
Sandboxing is really hard especially when facing our cyber capable models. So, we used our cyber capable model to red-team our own sandbox stack and found issues in many layers of the stack, including but not limited to Kata, gVisor, docker, and Linux.
The critical Brotli RCE vulnerability discovery came out of our large scale infrastructure code scanning efforts
Great research by @k_tr4n and @_sy1vi3 in assessing the widespread impact of this bug
Update your Node.js, Bun, Envoy, anything using Brotli ASAP!
Team work makes the dream work <3
We wanted to do something a bit different than using LLM's to generate all the pieces from exploit, writeup and art. We decided on a human created comic to offset the LLM usage in our release, that tells the tale with some exaggerated scenarios inspired from a few movies.
If you liked the art (or want your own exploit story/comic..) check out her work at ibiscraft.art
Largest credit goes to @lcfr_hax for the earliest recent discovery of the bug class.
This underlying engine issue was so wide it invited a lot of parallel discovery. As we coordinated our findings the scope of the problem continued to expand!
Let’s set the record straight: @lcfr_hax was the first to discover php2shell and develop a working exploit. See his August 5 tweet: x.com/lcfr_hax/status/208503…
His work convinced the PHP Foundation to invest in fixing this entire bug class. He then worked around the clock to review, test, and retest the patches.
php2shell is the most consequential vulnerability I have ever helped discover. I am proud of how our team worked closely in coordination with the PHP Foundation to provide support and to harden the patch in preparation for this release. Please read more about our approach:
When testing Astra on ExploitBench we found that it was able to achieve 100% ACE on all 41 CVEs
So we made a contamination free internal port using only V8 CVEs from the past 3 months. Astra still showed a major capability increase over 5.6 and did so with a lot fewer tokens!