Signal Scout
Measured 2026-09-09 · 251,174 repositories

The AI security gap

Agent development and the tooling meant to secure it are not growing together. One of them barely exists.

96%

of the 10,085 organizations building AI agents or MCP servers carry no AI security, red-teaming or evaluation tooling anywhere in their public code — 9,718 of them.

Building agents or MCP 71,128
Observability or evals 5,770
Runtime guardrails 333
Red teaming or scanning 22

Bars are to scale against the agent-building population, except that the two smallest are floored at a visible sliver so they read as a quantity rather than a rendering fault: at true scale they would be under one pixel. Only 208 of those 71,128 repos (0.29%) carry any security tooling at all.

Runtime guardrails, by vendor

Microsoft Presidio244
NeMo Guardrails53
Guardrails AI42
LLM Guard17
Rebuff1

Filters and policy enforcement that try to stop an attack while it happens. 333 distinct repos.

Red teaming and scanning, by vendor

garak10
PyRIT9
Giskard4
ModelScan1
Snyk1

Tools that attack or scan a system to find its weaknesses. Counted separately on purpose: merging them with guardrails would let offensive testing inflate a claim about defensive coverage. 22 distinct repos.

Notes