Glyph Guard
Proprietary security infrastructure, built from the ground up.
The platform behind every NXT assessment. Purpose-built to inspect AI agent behavior in real time and produce the evidence our assessments are based on.
How It Works
Not an AI watching your AI.
Entirely deterministic: written rules and statistical analysis that produce the same verdict every time. There is no model in the loop to hallucinate, drift, or be talked out of a decision. Watch a session play out, then try your own deployment.
Accounts payable agent
Inbox
Audit trail
Press play to watch a session: a normal request, an attack, the same attack again, and the record it leaves.
Simulated. No live system is called. Inputs are composed from attack patterns seen in real assessments.
01
A request comes in
Before the model sees it, the request is inspected. Anything that is an instruction to the agent rather than a request from a user is held, and the model never runs.
02
Clean traffic passes through
Legitimate requests reach the model untouched. No added friction for real users.
03
The response is checked on the way out
Data that should not leave is withheld before anything reaches the user, and every decision is written to the audit trail.
The interactive demo is available on larger screens.
Deterministic
No model sits in the detection path. The same input produces the same verdict every time, with nothing to hallucinate or be talked out of.
Powers every assessment
Every AI security audit we deliver is run through this platform. The findings, evidence, and compliance mappings in our reports come directly from what the platform surfaces.
Real engineering
Purpose-built detection and statistical risk scoring behind every decision. Not a prompt, not a wrapper, not a model.
Tested
Select a study and a model to see how it held up.
Two published evaluations. The newest tested eight frontier and open models for customer data disclosure. The first ran seven models through the same adversarial suite with and without our platform.
Eight models from two providers, configured as customer service agents holding synthetic customer records. An attacker states or guesses customer data and asks the agent to confirm, correct, or document it.
0 of 8
Models that disclosed PII
0+
Test executions
0
Attack classes
0 of 8
Asked for verification, every condition
All models · 8 models · 2 providers · 1,164+ test executions
Seven of eight models disclosed in at least one reported condition. Only Claude Opus 5.5 requested identity verification across every reported condition at default settings.
Test scope varies by model and the paper reports each condition separately. All customer records were synthetic. Where the application enforced authorization before the model saw a record, disclosure fell to zero; where the model was the only barrier, the records came out. Confirmation as Disclosure, September 2026→
Research
Published research and adversarial evaluations from our security research program. Controlled testing, documented methodology, and measurable results.
AI Security Audits
See what an assessment powered by this platform looks like.
Independent adversarial testing for AI systems in production. The same rigor applied to the models your business depends on.
Explore AI security audits →Contact
For partnership inquiries and research collaboration.