2026-10-15
Cybersecurity with local models
On CyberGym and ExploitBench the open-weights models have developed real offensive capability. Locally a SOC or red team uses them on confidential logs and code without leaking anything, and without a public API's refusals.
Offensive and defensive security work on material that can't leave the perimeter: proprietary code, incident logs, customer data. It's also the field where a public API refuses legitimate requests most often — an exploit analysis for an authorized penetration test, for example. Both problems go away when the model runs in-house.
Cyber capability that's now concrete
Z.ai writes it plainly in the GLM-5.3 card: cyber capability emerged faster than expected as post-training scaled, with the largest gain higher up the exploitation chain, where the model more than doubles GLM-5.2. GLM-5.3 is state of the art on CyberGym (84.5) for vulnerability discovery.
DeepSeek-V4.1-Flash beats it: 88.1 on CyberGym and 62.8 on SEC-Bench Pro in DeepSeek's table. MiMo-V2.6 folds cybersecurity directly into the RL stage: the Flash-RL reports 95.1 on CyberGym, and the Pro 80.2 on MiMo Cyber Bench.
Scores are vendor-reported with different harnesses and largely measure defensive analysis and vulnerability discovery, not real-world exploitation. Use them to orient.
Specialized, unfiltered models
For red-team work the catalog includes dedicated variants, all runnable both locally with Ollama and on cloud GPUs:
- Qwen3.8-27B-Uncensored-Cyber: a 27B dense model, runs entirely in VRAM on an RTX 5090. The card reports 100/100 on its set of 100 cyber-offensive prompts (regex refusal harness), without losing factuality or math.
- OrcaSAQ-2-Cyber-27B: another 27B cyber variant, on an RTX 5090 or locally with Ollama.
These models answer offensive-security questions that public APIs filter out. Responsibility for use and compliance with the law rests with whoever runs them: use only for authorized work.
Why local, not via API
- Sensitive material stays in. Logs, exploits, code and customer data don't pass through an outside provider. The SophyAI endpoint listens on
127.0.0.1. - No refusal that blocks the work. An authorized analysis shouldn't stop because a remote filter mistakes it for abuse.
- Long contexts. GLM-5.3, DeepSeek-V4.1-Flash and MiMo reach 1M tokens: a whole codebase or a long incident trace fits in one session.
- Cost by the hour. Triaging a repository or analyzing a dump are long processes: you pay for the GPU by time and stop when you're done.
The other scenarios: local AI agents and agentic coding, running locally. Full model overview: Frontier open-weights models, running locally.