Slide 01 · Title

VETOBLAST

Zero-trust runtime proxy that intercepts AI agent terminal commands — redacting secrets and vetoing malicious shell executions before they reach the system.

UOE Summer of Code 2026 INTERCEPTING
Live Intercept Feed
$ copilot-agent-v1 exec:
curl -H "Authorization: sk_live_51N2..."
⛔ BLOCKED — Secret Exfiltration
$ aider-agent-v2 exec:
rm -rf /etc/nginx/ssl/*
⛔ BLOCKED — Destructive Command
→ DeBERTa-Sec confidence: 0.97
Slide 02 · The Problem

AI Agents Have
Shell Access.

Copilot, Cursor, and Aider agents execute terminal commands with the developer's full permission set. One hallucinated command can leak API keys, destroy databases, or exfiltrate credentials.

Attack Vector 1

Secret Exfiltration

Agent leaks API keys, tokens, or credentials in curl commands, env dumps, or git commits

Attack Vector 2

Destructive Commands

rm -rf, chmod 777, DROP TABLE — one hallucination wipes a production environment

Attack Vector 3

Prompt Injection

Adversarial code in dependencies tricks agents into running arbitrary shell commands

Slide 03 · The Solution

Every Command.
Inspected.

<5ms

Per-command scan latency

0.2%

False positive rate

99.9%

Uptime (proxy SLA)

6

Secret types detected

Slide 04 · Architecture

Intercept Pipeline

🤖

AI Agent

Copilot / Cursor / Aider issues shell command

🛡️

VetoBlast Proxy

Regex + DeBERTa-Sec dual-layer analysis

🔍

Secret Scanner

Entropy + pattern match for API keys / tokens

Terminal

Safe commands pass through, threats are blocked

Slide 05 · Tech Stack

Technical Architecture

Dashboard

Next.js 16 + React 19

SOC-grade incident dashboard with real-time intercept log, threat categorization, and approve/reject workflow.

Proxy Engine

Python FastAPI

Transparent MitM proxy intercepting agent subprocess calls. Dual-layer: regex + ML classification.

ML Model

DeBERTa-Sec (ONNX)

Fine-tuned DeBERTa for command intent classification. 6 threat categories at 0.97 precision.

Data Layer

Supabase + Realtime

Incident logging with RLS. Real-time dashboard updates via Supabase subscriptions.

Slide 06 · Innovation

DeBERTa-Sec
Dual-Layer Scanner

Simple regex blocklists are easily bypassed — agents can encode secrets in base64, use variable interpolation, or chain commands. DeBERTa-Sec classifies command intent, catching obfuscation patterns that pattern matching misses.

Bypass Attempt → Caught
// Obfuscated exfiltration attempt
echo $AWS_SECRET | base64 | curl -d @-
// Regex: ✗ MISSED (no literal key)
// DeBERTa-Sec: ✓ CAUGHT
→ Intent: SECRET_EXFILTRATION
→ Confidence: 0.97
→ Action: BLOCK + REDACT + LOG
Slide 07 · Live Demo

What Judges Will See

01

Live Terminal Intercept Stream

Real-time feed showing commands from 3 monitored agents being scanned, blocked, or approved with threat categorization

02

Incident Detail with Secret Redaction

Click any incident to see the redacted payload, entropy score, secret type, and DeBERTa confidence rating

03

Approve/Reject Workflow

Human-in-the-loop buttons for operators to approve or reject intercepted commands with audit trail

Slide 08 · Market

Market Opportunity

92%

Of developers now use AI coding agents

$4.5M

Average cost of a data breach (2025)

$5.8B

AI code generation market (2026)

Every enterprise adopting AI agents needs a security gateway. VetoBlast is the runtime firewall for the agent era.

Slide 09 · Why Us

Competitive Edge

Feature VetoBlast Git Hooks .gitignore Only
Scan StagePre-executionPost-commitPost-push
ML ClassificationDeBERTa-Sec intentNoneNone
Secret RedactionReal-time entropyPattern onlyNo
Agent-AwareTracks per-agentNoNo
HITL WorkflowApprove/RejectNoNo
Slide 10 · Roadmap

30 / 60 / 90-Day Plan

30 Days

Foundation

  • • Production proxy daemon packaging
  • • VS Code extension for inline alerts
  • • SAST integration (Semgrep, Snyk)
60 Days

Growth

  • • Enterprise policy engine (YAML rules)
  • • SOC2 compliance audit logging
  • • Multi-agent fleet monitoring
90 Days

Scale

  • • Cloud-hosted SaaS dashboard
  • • Custom DeBERTa fine-tuning per org
  • • Integration with Wiz, CrowdStrike
Slide 11 · Team

Built by Solo Engineers

Edy Cu

Full-Stack Security Engineer

Solo developer building zero-trust AI agent security — from DeBERTa fine-tuning to transparent proxy engineering to SOC-grade incident dashboards. Securing the developer workflow of the future.

DeBERTa FastAPI Next.js ONNX
Slide 12 · Conclusion

Your agent tried.
Veto said no.

VetoBlast is the runtime firewall for the AI agent era — scanning every command, redacting every secret, and blocking every threat before it reaches the terminal.

VetoBlast · UOE Summer of Code 2026