LIVE FEED
ATLAS OWASP HIGH Significant risk · Prioritise patching RELEVANCE ▲ 6.2

Canary Tokens Expose AI Supply Chain Risks in OT Environments

TL;DR HIGH
  • What happened: Canary tokens deployed by Dragos reveal active AI supply chain threat activity targeting OT environments.
  • Who's at risk: Operators of critical infrastructure — including energy, water, and manufacturing sectors — integrating AI components without supply chain verification are most exposed.
  • Act now: Audit all AI/ML components integrated into OT environments for provenance and integrity · Deploy canary tokens within AI pipelines to detect unauthorised access or tampering · Establish a software bill of materials (SBOM) process covering AI model and dataset dependencies
Canary Tokens Expose AI Supply Chain Risks in OT Environments

Overview

Dragos, a leading operational technology (OT) cybersecurity firm, has published research examining AI supply chain vulnerabilities using canary token instrumentation. The research, surfaced via Hacker News and published on the Dragos blog, demonstrates how canary tokens — lightweight tripwire artefacts embedded within files, URLs, or credentials — can be used to detect unauthorised access or exfiltration attempts targeting AI components integrated into industrial environments.

As AI acceleration tools such as Dragos’s own EmberAI become embedded within OT platforms, the attack surface expands beyond traditional IT supply chains into environments that directly influence physical processes. The implications for critical infrastructure operators are significant.

Technical Analysis

Canary tokens function by generating unique, trackable identifiers embedded within documents, model files, API keys, or dataset artefacts. When an attacker accesses or exfiltrates a canary-embedded resource, a callback is triggered, alerting defenders to the breach.

In the context of AI supply chains, canary tokens can be embedded within:

  • Model weights or configuration files distributed via public or internal repositories
  • Dataset artefacts used for fine-tuning or RAG pipelines
  • API credentials bundled with AI SDKs or integration packages

Dragos’s research indicates that by seeding these tokens across the AI component supply chain within xOT (extended OT) environments, security teams can gain early warning of compromise — particularly relevant where AI models are sourced from third-party vendors or open-source repositories without rigorous vetting.

Framework Mapping

MITRE ATLAS:

  • AML.T0010 (AI Supply Chain Compromise): Directly applicable — adversaries targeting AI components before they reach OT deployment.
  • AML.T0115 (Publish Poisoned AI Artifacts): Risk of malicious actors distributing tampered model files through legitimate-appearing channels.
  • AML.T0109 (AI Supply Chain Rug Pull): Legitimate AI packages could be weaponised post-adoption.

OWASP LLM Top 10:

  • LLM05 (Supply Chain Vulnerabilities): The core risk category — third-party AI components lacking integrity verification.
  • LLM03 (Training Data Poisoning): Secondary risk where compromised datasets influence model behaviour in OT-integrated AI systems.

Impact Assessment

The sectors most directly at risk include energy, water, manufacturing, and transportation — all of which are increasingly adopting AI-driven analytics and automation within their OT stacks. A compromised AI component in these environments could enable:

  • Covert reconnaissance of industrial network topology
  • Manipulation of AI-driven process recommendations leading to unsafe physical states
  • Persistent access via backdoored model weights that survive routine security reviews

Nation-state threat actors with demonstrated interest in critical infrastructure disruption represent the highest-priority adversary profile for this threat vector.

Mitigation & Recommendations

  1. Deploy canary tokens throughout AI model and dataset pipelines to detect unauthorised access.
  2. Maintain a comprehensive SBOM for all AI/ML components, including model provenance, training data sources, and dependency chains.
  3. Cryptographically verify model weights and configuration files before deployment using hash validation or model signing schemes.
  4. Restrict AI component sourcing to vetted, internal repositories with change-control processes equivalent to those applied to OT firmware.
  5. Segment AI inference infrastructure from core OT control networks where feasible, limiting blast radius in the event of compromise.

References

◉ AI THREAT BRIEFING

Stay ahead of the threat.

Twice-weekly digest of critical AI security developments — every story mapped to MITRE ATLAS and OWASP LLM Top 10. Free.

No spam. Unsubscribe anytime.