split-brain-harness
Dual-hemisphere security layer wrapping any LLM: detects prompt injection, authority impersonation, and multi-turn escalation. Benchmarked on three adversarial datasets.
AI Security Research · North Shore, Oʻahu
Security infrastructure for
autonomous systems.
SGAIL Labs builds firewalls, detection models, and witness infrastructure for AI-to-AI systems — and turns that work into adversarial training datasets and new training methods other teams can build on. Conservative by design. Evidence at every layer.
Flagship
The SGAIL Firewall.
Our core product line: a deployable control point for AI-to-AI systems that inspects exchanges, enforces explicit policy, pauses high-risk activity, and preserves a tamper-evident chain of custody for review.
Enforcement · Witness · Evidence
Verify every exchange. Enforce policy at the boundary.
The SGAIL Firewall monitors agent-to-agent activity in real time, evaluates it against explicit rules, holds risky exchanges pending review, and writes a signed, hash-linked Witness log so every incident is reviewable and defensible.
- ✅ Real-time policy enforcement and execution pause
- ✅ Tamper-evident Merkle-chained Witness log
- ✅ Operator-controlled escalation and kill-switch support
- ✅ Offline / edge-first — conservative third-party surface
Status: Public OSS components + private alpha (accepting pilots)
Request a pilotSelected work
A published body of AI-security tooling.
Most of our detection stack is open source and published to crates.io and PyPI. These are the models the training datasets are built from.
deobfuscate
Multi-pass text deobfuscation and encoding-evasion detector — strips homoglyph, base64, Morse, and leet evasions before the model ever sees the input.
glyph-validator
Deterministic CJK glyph geometric-coherence validator — a data-integrity gateway for multilingual pipelines, no LLM in the loop.
Building datasets or new training methods?
SGAIL Labs is available for adversarial dataset construction, red-team corpora, and novel training/evaluation methods derived from our own detection models. Tell us what you're training.
Explore training work