Adaptive testing. Coverage-guided by default.
Blackbox tools guess from the outside and miss rare paths. Whitebox needs your source code. Verigrey is the sweet spot: an adaptive testing agent that sees your agent's trajectories, tool calls and reasoning — and gets smarter every run.
Blackbox misses the rare paths. Adaptive testing goes looking for them.
Blackbox eval tools
Only see inputs and outputs. Prompt-injection checklists and trial-and-error — no adaptivity, no depth, easy to fool.
Verigrey adaptive testing
Observes trajectories, tool selections, arguments and reasoning traces — guides exploration toward uncommon, high-risk states. No source access.
Whitebox analysis
Deepest visibility, but demands full source access and heavy engineering — impractical for third-party or fast-moving agents.
From an English sentence to a graded, replayable report.
Write the policy
State a guardrail as a plain sentence — "no PII after a tool call," "never refund without a matching order." No rule DSL to learn.
Formalize it
Verigrey converts your sentence into a formal, checkable test the agent must satisfy across its behavior.
Explore adaptively
The testing agent drives the target agent adversarially — observing trajectories, tool calls and reasoning to hunt uncommon states.
Return evidence
A graded, replayable report — with root cause and suggested fixes on any finding, mapped to your regulatory frameworks.
Every run teaches the next. Coverage-guided exploration means the next campaign starts smarter than the last — a compounding advantage manual red-teams cannot match.
Blackbox tools guess from outside. Verigrey sees the loop.
Sees trajectories
Every step your agent takes — not just its final answer.
Sees tool selection & args
Which tools your agent calls, with exactly what arguments.
Sees reasoning trace
Why the agent chose the action it chose.
Coverage-guided
Each run explores what previous runs missed.
Conversing with agent
An adversarial partner that drives the agent toward violations.
Root-cause + fix
Every finding is grounded — with a suggested fix, not just an alert.
Two decades of program analysis, pointed at agents.
Verigrey's engine descends from award-winning research in program analysis and automated program repair — the same lineage as SemFix and AutoCodeRover, whose techniques were acquired by SonarSource.
Autonomous agents are, at bottom, programs whose control flow is decided at runtime by a model. That is exactly the class of system our founding team has spent 25+ years learning to analyze: reasoning about vast spaces of possible executions, finding the rare path that breaks a property, and doing it automatically.
Applying coverage-guided exploration and automated program repair to agent behavior is what lets Verigrey find 9× more real vulnerabilities than expert manual testing — and do it in under an hour instead of weeks. Every finding is grounded in a concrete, replayable trajectory. Nothing to argue with.
Bring us a policy. We'll bring the proof.
Book a scoping session and we'll turn one of your real guardrails into a test plan.