IntentFence: AI Agent Workflow Security Copilot Meta Global AI Developer Hackathon 2026

Live Cloud Demo: https://agent-workflow-security.vercel.app

IntentFence is a denial-aware cybersecurity copilot designed to detect and block autonomous AI agents from evading access restrictions by switching tools, views, or delegating tasks after a sensitive operation is denied.

============================================================

Key Distinction: Stateless Policy vs. Objective-Aware Guard

1. Baseline Stateless Policy:
2. Evaluates each action in complete isolation.

3. Vulnerability: When an agent is denied access to a restricted resource (e.g., "payroll_records" via file_read), the agent immediately replans and calls a different tool (e.g., sql_query on "payroll_view"). The baseline check allows the second call because the tool and raw string differ, enabling unauthorized bypass.

4. IntentFence Objective-Aware Guard:

5. Binds actions to a stateful task and business objective.

6. Protection: When an action under an objective is denied, IntentFence permanently records a denial tombstone. Any subsequent action attempting to satisfy that same objective—whether through an alternative tool, alternative resource alias, or delegated sub-agent—is automatically blocked at the pre-execution boundary.

7. Granularity: Unrelated business objectives are unaffected and continue executing without false-positive lockouts.

============================================================

Live Comparison Example

Scenario: data/scenarios/alternative_route.json

============================================================

Current Platform Features

• Objective-Level Enforcement: Tracks agent actions across multi-step execution graphs.
• Correlation Engine: Automatically detects tool-switching and delegated child-agent evasions.
• Guarded Tool Proxy: Intercepts callable tools before execution, raising SecurityEnforcementError on blocked objectives.
• Interactive Security Sandbox: Evaluators and hackathon judges can enter custom objective IDs, agent IDs, and tools to test live pre-execution interception.
• One-Click Audit Export: Download complete SQLite telemetry directly as formatted JSON or CSV files from the web UI.
• Zero-Cost Local AI Copilot: Uses Meta Llama 3.2 (3B) via Ollama to generate structured SecOps advisory explanations. Enforcement logic remains strictly deterministic.
• Interactive Pastel Dashboard: Real-time telemetry streaming via Server-Sent Events (SSE) on port 8001.
• SQLite Persistence: Complete audit trails stored in data/intentfence.db (automatically adapts to /tmp on serverless hosts).
• Five Scenarios Included: Alternative route, tool switch, delegated agent, benign workflow, and repeated attempt.
• Cloud Serverless Ready: Native Vercel deployment support via vercel.json and api/index.py.

============================================================

Project Structure

• api/
  • index.py: Vercel serverless entrypoint for cloud deployment.
• vercel.json: Vercel deployment configuration.
• src/workflow_security/
  • events.py: Pydantic schemas for action events and security outcomes.
  • policy.py: Canonical resource mapping and baseline access policies.
  • guard.py: Objective-level stateful denial tracker.
  • correlator.py: Pattern matcher for tool-switching and delegation evasion.
  • ai_analyzer.py: Local Ollama interface for Meta Llama 3.2 (3B).
  • agent_adapter.py: GuardedToolProxy wrapping tools for runtime enforcement.
  • db.py: SQLite schema and persistence functions (with cloud serverless fallback).
  • web_app.py: FastAPI web application with SSE streaming, sandbox tester, and audit export.
  • simulator.py: Batch runner for synthetic scenario evaluation.
• tests/
  • test_workflow.py: Unit tests for core algorithms and models.
  • test_realtime_and_adapter.py: Integration tests for adapter, sandbox, export, and streaming endpoints.
• data/
  • intentfence.db: SQLite database of runs and security telemetry.
  • scenarios/: 5 preconfigured attack and benign workflow test fixtures.
• docs/
  • architecture.md: Full architectural specification.
  • project_guide.md: Comprehensive how-to-work guide.
• run_dashboard.py: Launcher for the interactive web dashboard on port 8001.
• requirements.txt: Python package dependencies.
• pytest.ini: Pytest configuration with pythonpath set to src.

============================================================

How to Work on the Project Locally

1. Prerequisites:
2. Python 3.10+ (tested on Python 3.14).

3. Ollama running locally with Meta Llama 3.2 (3B):

   • Start daemon: ollama serve
   • Pull model: ollama pull llama3.2:3b

4. Running Automated Tests:

5. Command: pytest

6. Result: 20 passed in under 3 seconds.

7. Running the Web Dashboard:

8. Command: python run_dashboard.py

9. Open browser: http://127.0.0.1:8001

10. Available features:

    • Top navigation bar (Dashboard, Developer Sandbox, Protection, Audit).
    • Scenario dropdown selector.
    • Live Server-Sent Events simulation stream.
    • Interactive Security Sandbox to test custom actions live.
    • One-Click Telemetry Export (JSON and CSV).
    • Real-time Meta Llama 3.2 copilot security advisory.

11. Using the Guarded Tool Proxy in Code:

    from workflow_security.agent_adapter import GuardedToolProxy
    from workflow_security.guard import WorkflowGuard
    from workflow_security.events import ActionEvent
    

guard = WorkflowGuard(restricted_resources=["payroll_records"]) proxy = GuardedToolProxy(guard=guard)

def my_file_tool(path: str): return f"reading {path}"

proxy.register_tool("file_reader", my_file_tool)

event = ActionEvent( task_id="task-01", objective_id="export-payroll", agent_id="agent-1", tool="file_reader", resource="payroll_records", status="REQUESTED" )

result = proxy.execute(event, path="payroll_records")

============================================================
Deploying to Vercel
============================================================

IntentFence is preconfigured for zero-configuration deployment to Vercel:

1. Import Repository:
- Go to https://vercel.com/new
- Select your GitHub repository: mdbaasim/agent-workflow-security

2. Deployment Settings:
- Framework Preset: Other
- Root Directory: ./
- Build Command: (leave empty)
- Output Directory: (leave empty)

3. Automatic Configuration:
- Vercel automatically detects vercel.json and builds api/index.py with @vercel/python.
- The SQLite persistence layer automatically detects the VERCEL environment variable and writes to /tmp/intentfence.db to comply with serverless storage requirements.

============================================================
License and Disclaimer
============================================================
Developed for the Meta Global AI Developer Hackathon 2026.
Educational and research prototype demonstrating defense-in-depth against autonomous agent evasion attacks.

Built With

• python