What's this about?
As models get better at code, their cybersecurity capability rises with it, and closed models currently lead. CyberFactory is about harvesting real-world security instances at scale to train on. 28 upvotes.
What exactly is the dual-use problem?
The skills are identical. Finding a vulnerability so you can patch it and finding one so you can exploit it are the same act; only intent differs. So every advance in defensive tooling is also an advance in offensive capability, and publishing a method to scale one scales both. That tension runs through the Trail of Bits piece and the tablet-rooting story from yesterday's digest as well.
What's the community saying?
Interest is in the pipeline rather than the results — where you source real security instances at scale, and whether models trained this way generalise past the specific bug classes they were fed. It arrives the same week as concrete evidence that these capabilities work in practice.