NanoMind security classifier

The same published model HackMyAgent loads, run without the pipeline around it
This is the smallest, fastest tier of NanoMind: a high-recall pre-filter (Mamba TME, 8.3 MB ONNX, runs on CPU in sub-millisecond time). It reads an artifact and returns a raw attack-pattern affinity across 10 classes. It is not a verdict, and it will not match what HackMyAgent reports. The raw tier over-flags ordinary prose, and it can also miss an attack the scanner catches: hidden-character payloads are handled in HackMyAgent by a separate deterministic analyzer that is not part of this model. Run npx hackmyagent secure for an actual finding.
Examples (the first two are benign and both over-flag; the rest are attacks)
Paste an artifact snippet or pick an example. The inline tier reads it and shows the raw 10-class attack-pattern affinity.

What this is. The open, on-device classifier tier that ships inside HackMyAgent. Model card and files: opena2a/nanomind-security-classifier.

What it is not. A standalone verdict. The 10-class affinity is one input to HackMyAgent's semantic compiler, which corrects benign context and escalates to the reasoning tier, the analyst, before producing findings.

The full NanoMind model line: collection: NanoMind AI agent security models (classifier, analyst, and the Metal-stable MLX analyst build).

Learn more at opena2a.org/nanomind. Apache-2.0. Classifier 0.5.0, 10 output classes: exfiltration, injection, privilege_escalation, persistence, credential_abuse, lateral_movement, social_engineering, policy_violation, benign, steganography.

On the steganography class. It is an output label of this model, not a hidden-character detector. Measured behavior: it fires on ordinary prose and misses zero-width, bidi-override and variation-selector payloads, because such a payload has no whitespace and collapses to a single unknown token. HackMyAgent detects those with a separate deterministic byte-level analyzer that this Space does not run.