Prompt injection incidents

Content the agent read (an email, web page or file) made it act for someone other than its user.

1incident documented across the ecosystem

Vendor

OpenAI (12)Anthropic (6)Meta (2)Google (1)Moonshot AI (1)Tencent (1)Multiple AI providers (unnamed) (1)Cairn (open-source project) (1)

Severity

Prompt injectionHighOff-task agentReportedBefore Oct 22, 2026

OpenAI Codex follows a planted AGENTS.md in a user's Downloads folder and tries to install a malicious npm package

OpenAI · Codex

A security firm stopped OpenAI's Codex mid-session as it tried to install a malicious npm package on a customer's computer. The instructions came from an AGENTS.md file sitting in the customer's Downloads folder, which Codex was running in.

Data: free for non-commercial use with attribution to AgentVet.ai (CC BY-NC 4.0). Download CSV