When Trust in AI Becomes an Attack Surface: Analyzing Prompt Injection, Model Theft, and Pipeline Poisoning

The report details EchoLeak (CVE-2025-32711), a high-severity zero-click prompt-injection allowing Microsoft 365 Copilot to execute hidden instructions from an email and exfiltrate sensitive data, and situates this as representative of broader AI threats: indirect and audio prompt injection, tool/skill poisoning in agent registries (ClawHub/ClawHavoc), supply-chain compromises (Trivy/LiteLLM), pipeline poisoning that persists across sessions, and endpoint malware (EvilAI, Atomic Stealer). It highlights that traditional security controls are largely blind to these attack channels and recommends defense-in-depth mitigations: least privilege for agents/tools, validating model outputs before actions, human-in-the-loop for irreversible operations, supply-chain provenance and signing, sandboxed model loading, and behavioral monitoring of model I/O.