OrcaCyber Zero 1.0, post-trained on GLM-5.3 for cyber work, hits 98% pass@1 on CyberGym
TheZachMueller · x · 2026-09-30
OrcaRouter launched OrcaCyber Zero 1.0, a GLM-5.3-based model post-trained for vulnerability research, exploit reproduction, red teaming, and autonomous cyber workflows. It scores 98.07% pass@1 (1,478/1,507) on CyberGym Level 1, a set of 1,507 real-world vulnerability-reproduction tasks from 188 open-source projects.
The model offers a 1M-token context window, native function calling, JSON outputs, and configurable reasoning effort via an OpenAI-compatible API at $3/$5 per 1M input/output tokens. Access is gated and intended for trusted security researchers and authorized testing. The post frames this as the start of a cyber-capability race among open models, following Anthropic's warnings about GLM's cyber capabilities.
More from Models
- Meta Muse Spark 1.3 hits #2 on ProgramBench, full eval data open-sourced — jyangballin · 2026-09-30
- AI's "OH MY GOD!" exclamations may actually help its reasoning — danintheory · 2026-09-30
- ChatGPT web no longer shows which model generated a response — Lust_OfVelvet · 2026-09-30
- Blogger: OpenAI models crush Claude Opus at scientific automation, not just benchmarks — VraserX · 2026-09-30
- Codex voice quietly loses unlimited Pro access, now burns credits for everyone — MakerOfToys · 2026-09-30
- Claude weekly usage limits are so slow users joke about biblical timescales — _Stocko_ · 2026-09-30