News
Open-Source Vuln-Hunting Model apex-flash-1: Near Claude Opus 5 Results at 1/30th the Cost
Cantina Security and Yeta Labs released apex-flash-1, an open-weight model purpose-built for vulnerability research. Based on Zhipu's GLM-5.3-Flash and RL-trained on 50 real vulnerability cases, it solved 40 of 60 internal test tasks — just 3 fewer than Claude Opus 5 High — at roughly one-thirtieth the cost. MIT-licensed on Hugging Face.
In early October, security firm Cantina Security and Yeta Labs released apex-flash-1. It’s an open-weight AI model trained specifically for vulnerability research — based on Zhipu Z.ai’s GLM-5.3-Flash, fine-tuned with GRPO reinforcement learning. Released under MIT on Hugging Face, free for anyone to download and deploy locally.
Cantina Security was founded by Harikrishnan Mulackal, previously of the Spearbit security research network. The company sells automated security software plus researcher-led security services. This release packages real cases accumulated from its own vuln research into a reusable model.
Training Data: 50 Real Vulns, 150 Tasks
The training data isn’t synthetic. Cantina took 50 real, already-disclosed vulnerability cases and organized them into 150 training tasks. Each case comes in 3 versions: white-box with guided source, less-hinted source, and black-box with only the runtime target. RL rollouts ran in the Codex agent framework, in environments close to real software.
Case mix: authorization, identity, and privilege-escalation flaws at 72%; accounting and numeric-precision issues at 18%; the rest time-validation, business-rule, and SSRF. The model is 321.3B parameters (MoE, 18B active), needing ~640GB VRAM for BF16 deployment.
Results: Near Claude Opus 5 High, 31x Cheaper
Cantina tested 60 tasks across 20 held-out vuln cases. Results:
- apex-flash-1: solved 40 (66.7%), cost ~$2.38
- GLM-5.3-Flash (pre-fine-tune): solved 36 (60.0%), cost ~$4.56
- Claude Opus 5 High: solved 43 (71.7%), cost ~$74.68
Opus solved just 3 more tasks at 31x the cost. Per solved task: ~$0.06 for apex-flash-1 versus ~$1.74 for Opus.
Two reminders. First, this is Cantina’s own test set — company-reported, not independently evaluated. Second, Cantina positions apex-flash-1 as an “executor”: best orchestrated by a larger model, handling execution work like reading code, calling tools, writing PoCs, and verifying vulns.
The Controversial Bit: The Abliterated Variant
Cantina also released an “abliterated” variant. It modifies refusal behavior — fewer refusals, full researcher control over model behavior. This version wasn’t separately evaluated.
That puts the dual-use debate on the table: an automated vuln-hunting model with fewer refusals is usable by attackers too. Cantina’s response: attackers never followed acceptable-use policies anyway — they can run open models locally and strip safeguards. Restricting AI’s cybersecurity capabilities mostly blocks the rule-following defenders, not attackers.
Who Can Use It
MIT means commercial use is free too. Weights run on vLLM, SGLang, or Transformers. The cost is hardware: BF16 needs multi-GPU nodes, though the community already has a 4-bit quantized version with far lower VRAM demands.
For security teams, this drops AI vuln-hunting costs from “tens of dollars a run” to “a few dollars a run.” A nice touch for Chinese developers: the base is Zhipu’s GLM-5.3-Flash — another battle-tested derivative of a domestic open model.
(Sources: MarkTechPost, October 4; Complete AI Training; RuntimeWire. Benchmark figures are Cantina’s self-reported numbers.)