Cantina Releases Apex-Flash-1, an Open Model Trained on Real Flaws for Security Research

Out of 60 test tasks, the model solves 66.7% of cases for $2.38, compared to 71.7% for Claude Opus 5 High but at more than $74.

The cybersecurity company Cantina, in association with Yeta Labs, has made public the weights of apex-flash-1, its first model dedicated to vulnerability research. It was post-trained on real flaws from its proprietary dataset, made up of paid vulnerabilities.

Almost as good, thirty times cheaper

Out of a set of 60 tasks kept away from training, apex-flash-1 achieves a 66.7% success rate on the first try for a total cost of $2.38. For comparison, GLM-5.3-Flash tops out at 60% for $4.56, and Claude Opus 5 in High mode goes up to 71.7%, but for $74.68.

Cantina assumes a bias in the debate on access to AI's offensive capabilities: according to it, restricting these tools to proprietary models mainly penalizes defenders, who are subject to purchasing and compliance rules, while attackers can already hijack open models. The company compares the situation to past controversies around Nmap or Metasploit, which have become basic tools of the trade. It announces that it will continue training with more complex exploitation chains and environments drawn from public CVEs.