Dell x NVIDIA Hackathon 2026

SquidWard

Enforce agentic traffic
at AI speed.

AI agents connect to dynamic endpoints faster than static firewall rules can keep up. SquidWard detects behavioral risk locally and turns it into analyst-approved agent policy.

TEAM V-X / ENTERPRISE AGENT SECURITY / 01
TEAM V-X / 02

Four builders. One local defense loop.

PROJECTSquidWard/TEAMV-X
THE SPEED GAP / 03

Traditional firewalls were built for humans and static traffic.

A

Human-speed
firewall policy

Static destinations, manual rule changes, and ticket-driven response.

Predictable traffic
VS
B

AI-speed
agent traffic

Autonomous actions, newly discovered tools, and dynamic endpoints.

Constant change
SquidWard closes the speed gap. It turns live agent behavior into safe, enforceable policy updates.
PRE-AI IT SYSTEMS / 04

Pre-AI IT Systems

LIVE TRAFFIC
Internal systemsUsers · laptops · applications
→requests
FirewallStatic rules · known destinations
→allowed traffic
InternetExternal destinations
POST-AI IT SYSTEMS / 05

Post-AI: SquidWard keeps the firewall current.

LIVE TRAFFIC
Internal systemsUsers · laptops · agents
→requests
FirewallSquid Proxy today · expandable
→allowed traffic
InternetExternal destinations
↕reads current rules
IT operatorsAsk · guide · review
→human input
SQUIDWARD / SECURE GB10 BOUNDARY
SquidWard agentOpenClaw · autonomous decisions
→
APIQuestions · policy tools
→
GB10 + local LLMPrivate reasoning
↕
Rules databasePrivate policy source of truth
PRIVATE BY DESIGN Rules, questions, and inference stay inside.
↔scheduled scan
Public CVE databaseExternal threat intelligence
DEMO / 06

Demo

ALWAYS-ON DETECTION / 07

Fast and slow detection catch different threats.

<1sLIVE PATH

Immediate anomaly scoring

Rules, rolling baselines, and CPU Isolation Forest score each new event without depending on an LLM.

HISTORYOFFLINE PATH

Sequence-level analysis

A safe SQLite snapshot and GPU PyTorch model find slow, related patterns across longer windows.

DETERMINISTIC EVIDENCE→ALWAYS-ON SECURITY AGENT
SAFE AUTONOMY / 08

The model recommends; a human authorizes; the agent enforces.

policy_recommendation {
  action_type: "deny_destination",
  target:      "test-storage.local",
  scope:       "business-agent",
  expires_at:  "2026-07-27T..."
}
H
HUMAN GATE
✓ENFORCEDAgent policy
No shell commandsNo direct enforcementSchema validatedFully audited
TECHNOLOGY STACK / 09

One GB10 runs the entire agent security stack.

04ANALYST
EXPERIENCE
WEB CONSOLEReact + ViteIncidents · approvals · audit
↔
APPLICATION APIsFastAPI + FastMCPREST · MCP tools · policy surface
03AGENT +
CONTROL
TRAFFIC PLANESquid Proxy 6.6Observe · allow · deny
→
PYTHON SERVICESIngest + Detect + StoreCollector · models · SQLite
→
AGENT RUNTIMEOpenClawInvestigate · explain · recommend
02LOCAL
INFERENCE
MODEL GATEWAYLiteLLMAuth + OpenAI-compatible API
→
GPU SERVINGvLLMHigh-throughput inference
→
LOCAL MODELQwen3.6-27B-FP8Reasoning + tool use
01COMPUTE
FOUNDATION
Dell Pro Max with GB10NVIDIA GB10 · CUDA 13 · Docker · 128GB unified memory
100%LOCAL + PRIVATE
STATS FOR TODAY / 10

Today, local inference delivered real usage at zero model spend.

01 / CLOUD3 TEAM REPORTS
$526.24spent today
800.2Mtokens used
14models used
CACHE HIT94.9–100%
02 / LOCAL GB10PRIVATE INFERENCE
$0.00model spend
187total requests
167successful
3.38Mtokens used
MODELQwen3.6-27B-FP8
SQUIDWARD / 11

Agentic traffic needs a firewall
that can keep up.

Observe continuously. Adapt at AI speed. Enforce with control.

LOCAL-FIRSTBUSINESS VALUEEND-TO-END DEMOTECHNICAL CONTROL
SPEAKER NOTES