Ghost in the Pipeline: Hunting Supply Chain Attacks from npm to AI Agents
Vulnerability Engineer @ Zoho
View Profile
Talk Abstract
Most supply chain detection assumes the payload is still on disk when you go looking for it. The trojanized axios packages broke that assumption. The dropped RAT cleaned up after itself, and during the incident response a meaningful portion of confirmed-compromised endpoints came back perfectly clean to every scanner pointed at them. The only reason those machines were caught at all was passive DNS telemetry — the network proof that they had reached out to the C2 domain `sfrclak.com`, even though there was zero filesystem evidence left.
That gap — between "this machine was compromised" and "it can't be proven with endpoint tooling" — is what forced a rethink of how supply chain detection works at fleet scale. Looking at AI/ML pipelines through the same lens, the attack surface is wider, the blast radius is larger, and the detection tooling barely exists.
About the Speaker
Harish Ravichandra
Vulnerability Engineer @ Zoho
Harish Ravichandra is a Security Engineer specializing in endpoint security operations, vulnerability management, and supply chain threat detection at enterprise scale. His core operational stack includes CrowdStrike Falcon (sensor deployment, Falcon Data Replicator for passive DNS and process telemetry, and Real Time Response for live endpoint forensics), Tenable.io for continuous vulnerability assessment, and privileged access management across mixed OS environments (macOS, Linux, and Windows).
He writes detection and remediation scripts for emerging CVEs, supply chain compromises, and zero-day threats across macOS, Linux, and Windows, covering the full lifecycle from initial triage and threat intelligence correlation to fleet-wide deployment and post-deployment validation. His operational scope includes software supply chain attacks (npm, PyPI, GitHub Actions), endpoint hardening, vulnerability prioritization, and cross-platform forensic analysis using EDR telemetry.
He is the architect of an internal vulnerability intelligence platform that automates the full detection lifecycle, including CVE ingestion from NVD and vendor advisories, threat intelligence correlation against active exploit data, AI-driven detection script generation with platform-specific constraints, and orchestrated deployment across the endpoint fleet. The platform enforces a read-only detection scope, where AI-generated scripts can observe and report but cannot modify or remediate without explicit human authorization.
His current research focuses on extending supply chain detection methodologies from traditional package ecosystems into AI/ML supply chains, including model integrity verification, pickle deserialization attack detection, MCP tool poisoning, and agentic AI framework security mapped to MITRE ATLAS and OWASP LLM Top 10 2025.
View Profile