Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

294 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Agent Hypervisor

Deterministic virtualization of reality for AI agents.

Status: Research proof-of-concept. Stage 3 (Beta Product). Not a product.
Author: Personal project — does not represent Radware's position.


The Core Idea

AI agent vulnerabilities are not bugs. They are architecturally predictable consequences of agents operating with unmediated access to inputs, memory, and tools.

The standard response is behavioral: detect bad actions, filter bad inputs, block bad outputs. All probabilistic. All bypassable.

Agent Hypervisor asks a different question:

"Does this action exist in the agent's universe?"

Not "is it forbidden?" — but "does it exist?"

The agent never sees raw reality. It sees a virtualized world defined by a World Manifest — a compiled specification of what actions exist, what trust levels grant what capabilities, and what data can flow where. Dangerous actions are not prohibited. They are absent.


Getting Started

5–10 minutes to first result: docs/quickstart.md

Start the gateway, open the Web UI, observe an attack blocked in real time, and change a manifest rule to see how it affects enforcement — no API keys required.


Proof Artifacts

The articles in this series are backed by runnable code. Each claim maps to a specific executable artifact.

Article Core claim Executable proof
1 — Every AI Defense Broke Permission security fails by design python examples/poisoned_tool_output_demo.py — baseline attack succeeds; hypervisor blocks it
2 — AI Aikido Stochastic design-time → deterministic runtime awc run --scenario unsafe --compare — shows raw surface vs. compiled boundary
3 — Design-Time HITL O(n) runtime HITL doesn't scale; O(log n) design-time does python _research/benchmarks/replay.py --walkthrough — Design→Compile→Deploy→Learn→Redesign cycle
4 — MCP and the Missing Layer Tool virtualization breaks the attack chain docker compose up gatewayhttp://localhost:8090/ui — live gateway with provenance firewall

Benchmark result

AgentDojo workspace benchmark (560 task × attack pairs):

Metric Value
Attack success rate (ASR) 0.0% — all attacks contained
Utility (safe task completion) 80.0% — false-deny rate near zero
Policy evaluation latency ~0.5 ms per call

Run it yourself: python _research/benchmarks/run_scenarios.py
Verify determinism: python _research/benchmarks/replay.py


Architecture

[ Raw Reality ]
      ↓
┌─────────────────────────────────────┐
│  Layer 0: Execution Physics         │  Container / network isolation
│  Layer 1: Base Ontology             │  What actions exist (design-time)
│  Layer 2: Dynamic Ontology          │  What the agent can propose now
│  Layer 3: Execution Governance      │  Allow / Deny / Ask / Simulate
└─────────────────────────────────────┘
      ↓
[ Agent — virtualized world ]

Manifest Resolution Law:

proposed action
  ├── explicit allow in manifest     → ALLOW
  ├── explicit deny in manifest      → DENY
  ├── invariant violation            → DENY
  └── not covered by manifest
        ├── interactive mode         → ASK
        └── background mode         → DENY

The world is closed-for-execution, open-for-extension.

Full architecture: WHITEPAPER.md


Key Documents

Document What it is
docs/quickstart.md Start here — 5-10 min walkthrough
WHITEPAPER.md Full architecture: four-layer model, AI Aikido, World Manifest Compiler, Design-Time HITL
docs/architecture.md Runtime and compilation paths; component map
scenarios/zombie-agent/SCENARIO.md ZombieAgent attack and how AH breaks it
manifests/example_world.yaml World Manifest template
manifests/schema_v2.yaml Full v2 schema reference

Runnable Demos

# Poisoned tool output: attack succeeds without hypervisor, blocked with it
python examples/poisoned_tool_output_demo.py

# Scenario suite (9 scenarios: attack / safe / ambiguous)
python _research/benchmarks/run_scenarios.py

# Trace replay — verify determinism, walkthrough the design cycle
python _research/benchmarks/replay.py --walkthrough

# Web UI gateway (requires Docker)
docker compose up gateway
# then open http://localhost:8090/ui

The Key Distinction from CaMeL

CaMeL (Google DeepMind, 2025) shares the same foundations: capability-based security, information flow control, a protective layer around the LLM without modifying it.

The architectural difference is when the LLM operates:

CaMeL Agent Hypervisor
LLM role in enforcement Extracts control flow at runtime Generates policy artifacts at design-time
Runtime enforcement LLM on critical path Deterministic lookup tables only
Policy scope Per-query Per-workflow (World Manifest)
Cross-session taint Not addressed Core scenario (ZombieAgent)

Honest Constraints

This is bounded, measurable security — not perfect security.

  • The World Manifest covers what was anticipated at design-time. Novel attacks require redesign.
  • Semantic ambiguity ("forward this to Alex") is not resolved — it is the open "semantic gap" problem.
  • Manifest authoring tooling (AI Aikido pipeline) is not yet implemented.
  • The 0% ASR result is on a specific benchmark with specific attack patterns. Not a universal claim.

Personal research project. Does not represent Radware's position.
References are to published research only.

About

No description, website, or topics provided.

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages