MITRE ATLAS
v5.4.0Real-world ATLAS case studies — follow a technique to its detail.
AML.CS0005 1 techniques
GPT-2 Training Data Extraction
Researchers demonstrated extracting verbatim training data through targeted probing.
AML.CS0010 1 techniques
Tay — Adversarial Training
Microsoft Tay was poisoned through user inputs that shifted its conversational model toward offensive content.
AML.CS0017 1 techniques
Bing Chat Indirect Prompt Injection
Hidden text on a webpage redirected Bing Chat to follow attacker instructions when summarising the page.
AML.CS0019 1 techniques
Bard Prompt Injection via Email Summary
Adversarial email body redirected Bard to follow attacker directives when summarising.
AML.CS0024 3 techniques
ChatGPT Plugin Data Exfiltration
Demonstrated indirect prompt injection via crafted documents that hijacked a ChatGPT plugin to exfiltrate chat history.