Martyx00

My project had two essential parts. In the first part, I used Meta models to generate skills based on rather detailed instructions. The second part was putting these skills through a real-life test. For the second part, I used not only Meta models but also GLM 5.3 and Claude Opus 5. I also verified the skills with multiple harnesses, not just Muse Code. report-triage skill: I took three examples of known CVEs, crafted vague reports for them, and tested whether the skill would help me identify the exact spot in the code. It needed some fine-tuning, but in the end it proved to work. sbom-recovery skill: I used my ancient ESP32 project for monitoring my hedgehog (movement, temperature, an SPI camera, and 3 different ways of connectivity, all bundled in a folder as a huge pile of mess) as a test run for it. stride-design-review skill: Untested during the hackathon. I will get to it in the coming weeks. code-security-review skill: Tested on known CVEs in the mbedtls and curl repositories. Surprisingly good accuracy on the first try, but rather slow, and it burns quite a lot of tokens. This one will need some fine-tuning in the coming weeks. binary-vuln-research skill: Tested on the wireless driver blobs of ESP32 and Realtek MCUs, trying to identify new vulnerabilities by giving the model access to Ghidra MCP, three dev boards from each vendor, and a HackRF knock-off with poor RF properties. The results were quite amazing: TODO new vulnerabilities were found in the latest versions of the drivers. Of these, 6 are confirmed RCEs. TODO have already been reported to the vendors; TODO are still pending manual confirmation and will be sent in the coming days. Also pending is testing Muse Spark 1.3 and Muse Wonder Full for these cases, and combining all these models in a meaningful way when using these skills. Please note that I cannot share the video via YouTube and I would also like to ask you not to share it further or make it public as it contains a capture of building PoC for an unfixed RCE vulnerability in ESP32 ZigBee drivers.