Why the ChatGPT Leak Proves We Need a Zero-Trust Harness for AI Agents
If your AI agent gets tricked by a prompt injection, a firewall won't stop the data leak—but a cryptographic kill switch will.
At Black Hat 2026, our setup at the Atsign booth was about as agile as it gets. Armed with a flexible display, high-tech demos, and a literal stack of velcroed banners that allowed us to tweak our messaging on the fly, co-founder and CTO perspective in hand, we set out to test our core thesis against the sharpest minds in cybersecurity:
Organizations don’t need more firewalls, complex VPNs, and open ports—they need a fundamental shift in how people, entities, and things communicate.
By day three, we had all the proof we needed. From ethical hackers trying to compromise microcontrollers to enterprise executives taking our cards to hand off to tech giants, the message was clear: the industry is actively searching for what Atsign has built.
To show rather than tell, we worked with Intel and ported their open-source Smart Route Planning Agent (a demonstration of Smart City traffic routing), and ran it unmodified using the Atsign platform rather than APIs and open ports for communication between the component parts.
A Linux machine runs the route planner and its services; two Android tablets simulate a roadside sensor and the other runs a commuter's app.
The tablet’s camera detects vehicles on-device (we used a toy car), publishes an encrypted traffic report, and the planner reroutes and pushes the new route to the commuter app, all with no open ports, no shared credentials, and all over an open WiFi network.
When booth visitors doubted how quickly code could be migrated onto our platform, we walked them through the exact stack we used:
The entire port took a single afternoon.
To show this isn’t just for core devs, we invited Atsign customer Rick Deacon to join us in the booth. A practitioner rather than a full time developer, Rick used Atsign AI Architect to build NeuroVitals—a secure, privacy-first mental healthcare app—entirely on his own, way ahead of schedule.
To put our software-defined perimeter to a stress test, we ran a three-day challenge featuring two ESP32 microcontrollers with displays (CYD S3 for the nerdy) connected to an unprotected Wi-Fi network, offering a mechanical gaming keyboard to anyone who could compromise them both.
Over three days, more than 35 unique MAC addresses launched continuous attacks against the devices. The result? Zero successful exploits.
Live from Black Hat 2026: Putting Atsign's zero open listening ports architecture to the test.
Bogdan Barchuk, a well-known ethical hacker who thoroughly probed the setup continuously for the three days, concluded:
"I see no potential for a successful attack beyond perhaps a denial of service."
For his methodical analysis and effort, he walked away with the keyboard.
The feedback throughout the week highlighted a growing shift in the market:
My personal highlight of the week came outside the convention center. Packed into a hotel elevator with a dozen Black Hat attendees, me carrying the demo equipment to my room, someone asked me for my elevator pitch. I had to keep it simple as I was getting out at the next floor:
"We allow people, entities, and things to communicate safely over hostile networks."
It earned a round of applause in the elevator, not what I was expecting but a happy surprise.
Black Hat 2026 proved that the industry is ready to move beyond traditional network security. Giving every actor on a network a cryptographic identity and eliminating exposed attack surfaces isn't just an idea, it's practical, and deployable today, and built for today's and tomorrow's autonomous systems and edge computing.
The firmware for our ESP32 challenge will be publicly available shortly for anyone who wants to run their own tests locally.