We train frontier-class models for offensive security, publish how we built them, and hand you the weights.

Model releases

Frontier models get bigger.
Ours get better.

Frontier models keep growing as their gains stall. Ours keep improving and stay small.

Performance on the frontier. Our models climb to match the frontier at a fraction of the size.
Powerful at a local size. Frontier models only grow larger. Ours hold their performance and grow with your hardware.
Vulnerability detection benchmark. Score is the percentage of samples where the model found the vulnerability and named its correct CWE. Run in the vScan harness on 1k samples across 20 languages, 23 June 2026.
Benchmark results compared with size. Top left is best. Score is the same found-plus-correct-CWE percentage. Some model sizes are estimates from public information.

Open models. Open harness.

  1. Speed
  2. Traces
Speed

Faster than the cloud. VAKT runs at over 13,000 tokens per second on an M5 Pro Mac.

Traces

Observable by default. vScan Studio saves the full chain of thought and the raw output of every tool call in your traces.

Rented model

  • Writes working exploits: Refused
  • Attacks live targets: Refused
  • AI sovereignty: Off
  • Uses your data for product improvement: On
  • Retains your prompts and logs: On
  • Trains on your IP: On

Your weights

  • Writes working exploits: On
  • Attacks live targets: On
  • AI sovereignty: On
  • Keeps your data on your servers: On
  • You set retention and audit logs: On
  • Bring your own provider: On

Rented model

  • Claude Code MCP serversSkillsChatsPluginsTools
  • Codex MCP serversSkillsChatsPluginsTools
  • Mistral Vibe MCP serversSkillsChatsPluginsTools

Switch platforms and you rebuild from zero

vScan Studio

  • Anthropic
  • OpenAI
  • Vaktex Cloud
  • Mistral

Provider broken? Bring your data and switch to the next

Open to every model.

Run any provider, any open model, or your own weights.

enesfrdenbsvfiplarhe