lupingqaq/inkblade is an agentic penetration testing framework for LLM training and inference clusters that use Kubernetes and GPUs. It is a Python repository, and its latest release is probe-v0.1.1. The provided information does not separately list anything that changed in this release, so the version can only be mentioned as the latest release that can be confirmed.
The tool works by placing static Rust probes inside the cluster and inspecting several infrastructure layers with declarative templates. The README explains that it checks scope, timing and approval records in approved environments, and collects evidence to build re-runnable rule bundles. FACTS lists 10 detection layers, 286 templates and 306 modules. The items include checks specific to machine learning infrastructure, such as trusted remote code execution, prefix cache separation and attestation chains. Results are exported as bilingual reports and reproducible rule packs, which is its defining feature.
However, since the README covers penetration testing and defensive bypass validation, it should be used only on assets with explicit authorization. The license is listed as NOASSERTION, so usage terms cannot be confirmed from this information alone. It has 2 stars and a momentum score of 1.7/10. On that evidence alone it is hard to call a widely adopted tool, and it is more reasonable to see it as a project focused on security checks for approved training and inference clusters.