HPE Fighting AI with AI: Defending Against Zero-Day Threats

Pradeep Hattiangadi and Naveen Toopran of HPE Networking demonstrate the adaptive threat defense framework for SRX firewalls, driven entirely from a third-party AI assistant. Using skills built on the Junos MCP server, the Juniper Secure Analytics MCP server, and Security Director Cloud APIs, they validate policy, tie AI application activity to user identity, and automatically contain a data exfiltration attempt. Hattiangadi, a technical marketing engineer for the SRX platforms, and Toopran, a product manager for SASE and security, presented at Security Field Day 16.

Hattiangadi played a SecOps administrator at a fictitious firm whose HR group was suspected of being targeted. From an assistant such as GitHub Copilot, he checked connectivity to the firewall, Security Director Cloud, and JSA. He then validated the HR group’s policy posture, found a missing enforcement rule, and applied the fix with a single confirmation. Security Director Cloud recognizes changes pushed directly to the firewall and does not overwrite them. Under the framework, users start on a baseline policy and HR uploads to sanctioned AI apps are monitored by new SRX data loss prevention capabilities. Repeated sensitive uploads within a set window raise a user’s risk and add controls. A detected reverse-connection exfiltration attempt moves the user into quarantine, with selective access left open so IT can fix the laptop. In the demo, one recruiter worked normally while Toopran found ChatGPT blocked. The assistant then built a timeline from logs filtered to each user’s IP, which showed the high-risk uploads and the threat that led to quarantine. A colleague on vacation showed no activity, which confirmed that nobody had borrowed her identity.

The team said the skills will be open sourced and work with any agent. The demo itself ran on an open-source model, while HPE fine-tunes its own back-end agents. DLP covers chat prompts, uploads, and downloads, and policies can be as granular as allowing uploads only to specific sanctioned apps. Security intelligence comes from ATP Cloud, and escalation feeds expire on a configurable timer, after which users return to normal access. A delegate argued that blocked users need a clear explanation, and the team pointed to a customizable block page. They framed the result as one closed loop in which an AI assistant configures, validates, investigates, and mitigates.