Resources & Insights
Trent Holmes
2 articles
-
ResearchScarecrow: Making the Attacker’s AI Say No
We developed Scarecrow, a tool that hides refusal instructions in documents so AI agents ingesting them are prompted to stop processing. Our findings show that this same technique can help defenders disrupt AI-driven access to sensitive documents.
October 5th, 2026TrendAI™ Research, Aya Debbagh, Trent Holmes
Read article -
BlogExploiting DeepSeek-R1: Breaking Down Chain of Thought Security
This entry explores how the Chain of Thought reasoning in the DeepSeek-R1 AI model can be susceptible to prompt attacks, insecure output generation, and sensitive data theft.
March 4th, 2025 5 minTrent Holmes, Willem Gooderham
Read article