Glyph Guard
The control layer between AI and action.
Runtime oversight for AI agents in production. Every request inspected before it reaches the model. Every response scanned before it reaches your users.
How It Works
Glyph Guard is not an AI watching your AI.
It's deterministic software: 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.
inspects every request, blocks threats in milliseconds
scans every response, stops data leakage
inspects every request, blocks threats in milliseconds
Modelscans every response, stops data leakage
Your UsersBefore the model ever sees it, Glyph Guard inspects it. Anything malicious or outside your policy is blocked in milliseconds, and the model never runs.
Inspection
<10ms
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.
Developer-first
A drop-in proxy, an SDK, and a CLI. Author policies as code, test payloads locally, and watch decisions stream to your terminal.
Real engineering
Twelve detector categories, six statistical risk engines, and a test suite behind every decision. Not a prompt, not a wrapper, not a model.
Integration
Sits between your agent and production. No retraining, no prompt changes, no model swap. Adopt it however fits your stack.
No code change to your agent. Point your existing OpenAI or Anthropic client at the proxy: every request is inspected in, every response scanned out.
# route through the guard, your key rides in a headerclient = OpenAI(base_url="https://guard.your-domain.com/v1",api_key="glyph", # placeholderdefault_headers={"X-Glyph-Api-Key": OPENAI_KEY,"X-Glyph-Agent": "support-bot",},)# your agent code is unchanged. blocked# requests never reach the model.
Use Cases
If an AI agent can touch real data or real systems, it can be attacked. These are the deployments where Glyph Guard matters most.
- -Block manipulation attempts before the bot ever reads them
- -Stop an agent from revealing one customer's data to another
- -Detect pressure that escalates gradually across a conversation
- -Keep every conversation inside the policies you define
- -Block manipulation attempts before the bot ever reads them
- -Stop an agent from revealing one customer's data to another
- -Detect pressure that escalates gradually across a conversation
- -Keep every conversation inside the policies you define
Capabilities
Full visibility into every request and response passing through your AI agents. See what your models are doing in production, not what you hope they're doing.
The Console
A system view, not a dashboard.
Security scoring, threat detection across twelve categories, six quantitative risk engines, and compliance coverage. Every decision, as it happens.
Security score
Live threat feed
Detection · 12 categories
Quant risk signals · 6 engines
| Agent | Trust | EWMA | CUSUM | SPRT |
|---|---|---|---|---|
| agent-sentinel | 92 | |||
| agent-codex | 78 | |||
| agent-nexus | 61 |
Compliance coverage
Mapped to OWASP LLM Top 10 and MITRE ATLAS.
Representative view. Live data streams from your own deployment.
Tested
Select a model to see how it held up.
Seven models, from a commercial API down to small models on edge hardware, each put through the same adversarial suite.
0
Payloads tested
0.0%
Stopped with Glyph Guard
0.0%
Model's own defense
+0.0%
Guard uplift
All models · 7 models · API + edge · 1,632 payloads · Phase 1 + Session 49, April 2026
Each model was run twice against the same payloads: once with Glyph Guard, once without. Detection includes the quantitative risk layer, which holds flagged responses before delivery; deterministic scanners alone average 96.0%.
Common Questions
The check itself is deterministic code, not a second model call, so evaluation runs in a few milliseconds. In SDK mode the engine runs in-process, so the overhead is near zero. In proxy mode the extra network hop puts measured end-to-end overhead around 100 to 200 milliseconds, which is small and predictable next to typical LLM response times of 500ms to 5 seconds. And blocked requests return faster than normal ones, since the model never runs at all.
Research
Published research and adversarial evaluations from our security research program. Controlled testing, documented methodology, and measurable results.
Contact
For demo requests, integration questions, or partnership inquiries.