Thanh M. Truong-Le
University of Information Technology, Vietnam
Titles as the sources gave them. In 2 entries, with 2 checked quotations.
What they said
Newest first. Each quotation was found word for word in the source.
“Safeguard coverage remains uneven: only 278 of 660 relationships are directly assessed, and strong coverage is uncommon.”
“Our analysis shows that tool, MCP, and action-layer attacks account for the largest share of atomic attacks, while the prompt-injection and action-hijacking class has the broadest paper coverage.”
In the Base
Every entry they appear in.
Agents of Chaos: A Systematization of Knowledge (SoK) on Malicious Attacks on Agentic AI Systems
A team from Vietnamese universities and Jiffy Labs reviews 98 sources on malicious attacks on AI agents, coding 284 distinct attacks and 452 safeguards into eight classes. They report that defences are unevenly assessed: only 278 of 660 attack-safeguard links were directly tested.
The Devil Wears Markdown: Detecting Malicious Intent Hidden in AI Agent Skill Instructions through Hybrid Static, Dynamic, and Agentic Analysis
According to the abstract, Thanh M. Truong-Le and colleagues test three open-source scanners on 140 agent skills and then propose a framework combining static, sandbox and LLM analysis, raising F1 from 0.70 to 0.85 when dynamic evidence is added.