Available for engagement
SGAIL Labs builds and curates adversarial security datasets derived from our own detection models and harnesses — and the novel training and evaluation methods that go with them. If you are training a model that has to survive hostile input, this is what we do.
Most security datasets are scraped or hand-labeled once and go stale. Ours are generated and verified by shipping detection systems: the deobfuscation engine, the split-brain harness, the unicode-interference and glyph-integrity validators. Every sample carries a provenance record and a machine verdict, so the labels are reproducible rather than anecdotal.
Dataset families
Adversarial instruction-override, authority-impersonation, and multi-turn escalation samples, graded by the split-brain harness across single- and multi-turn contexts.
Homoglyph, base64, Morse, zero-width, and leet transformations paired with clean originals — labeled by the deobfuscate engine for round-trip recoverability.
Text hidden in images — white-on-white, sub-pixel fonts, alpha overlays, EXIF, QR — extracted and scored by deobfuscate-vision. For training and stress-testing vision-language models.
Malformed, spoofed, and coherence-broken CJK glyphs from the glyph-validator, for multilingual data-integrity and tokenizer-robustness work.
Built to your threat model, your domain, and your target model — with a documented generation pipeline and a held-out verification split.
Methods
Beyond data, we develop the methods that make it useful — approaches that come out of our own security research.
Dual-hemisphere reconciliation: two independent lenses reach a verdict and a superposition-collapse probe resolves disagreement — a training and evaluation pattern for robust security classifiers.
Refusals and interventions graded against a tamper-evident witness log, so a model's safety behavior can be measured against a defensible ground-truth record rather than a static rubric.
Every training sample is tied to the detector, version, and verdict that produced it. Reproducible labels, auditable pipelines, and clean train/verify separation.
Whether you need an off-the-shelf adversarial split or a bespoke corpus and method for a specific model, we can scope it. One short intake gets it moving.
Start a dataset intake