Show HN: Beating GPT5.5-xhigh for Coding agent security with SLMs and IRM

A cybersecurity small language model has been trained and modified to outperform GPT5.5-xhigh on certain benchmarks. The model's reasoning was altered and its controls were supplemented using program analysis techniques like inline reference monitoring. This approach allowed it to achieve better results on challenging benchmarks such as LinuxArena and SleightBench. A free product based on this model is available at harden.run, along with full benchmark details in a blog post. The model is intended to secure coding agents, which can generate arbitrary code and pose complex security challenges.