The model reads a coding agent’s plan and selects project-specific security guidance before code is written.
Argos is a fine-tuned Qwen3.5-9B model. Corridor says it trained the model using thousands of synthetic and internal traces, with Baseten Loops for managed fine-tuning and reinforcement learning. The post does not state a standalone price, public weights release, or separate access terms for Argos.
On Corridor’s SecReT benchmark, Argos scored 0.655 micro F1 across 272 requests and 1,910 plan-item pairs. Corridor reports 77.8% recall, 56.5% precision, and a 740 ms mean time to first answer token; the base Qwen3.5-9B scored 0.480 micro F1.
In a SusVibes experiment using GPT-6 Sol to implement changes, Corridor says plan-selected security context raised security correctness from 32.8% with random context items to 53.2%. These are Corridor’s reported evaluations, not independent evidence that the model improves security across repositories or coding agents.