

LEVIATHAN
Agentic Workflow Interrogation
Investigate agent workflows. Run experiments. Verify security findings.
Capabilities
Put agent workflows to the test.
Sample Investigation
Approved A. Wrote B.
Two authorized callers shared a context. The owner’s approval for A triggered a write to B.
| Experiment | Receiver Result |
|---|---|
| A-Only Control | A Committed |
| Interleaved A/B | B Committed |
| Guarded B | B Denied |
3 / 3 hosted runs committed B.
Proven at the receiver.
Independent attempt logs and owner readback confirmed the B writes.
- Hosted Runtime
- ADK 2.13.0 · A2A 1.3.0
- Fresh Contexts
- 3 hosted runs · 3 B commits
- Source Replay
- Pinned ADK 2.12.0 · 3 fresh runs
One ID. Two invocations.
ADK reused a confirmation identity across A and B. The owner’s original response resumed B.
Local repair: 3 / 3 reattacks blocked B.
An ambiguity check on the pinned source preserved the authorized A control.
Bind approval to the exact invocation and arguments.
Investigate
Explore unfamiliar
agent software and
workflows.
Experiment
Build and run targeted
security experiments.
Verify
Validate evidence and
carry knowledge
forward.
Research Modes
Choose your depth.
Rapid Scan
Focused research on
a defined target.
Deep Scan
Extended investigation
of agent workflows.
Outputs
Evidence you can act on.
Review findings, experiments and supporting evidence.
Verified
Findings
Security findings checked
against evidence.
Experiment
Records
Tests and observations
behind each result.
Research
Context
Knowledge carried into
the next investigation.
How It Works
From scope to findings.
- Set Scope
- Choose a target and research objective.
- Investigate
- Run experiments with Leviathan.
- Review Evidence
- Inspect findings and plan the next investigation.