Hey CheckMates!
One of the themes that kept coming up in the last few weeks is how quickly indirect attacks are growing.
These look harmless at first:
- “Can you summarize this link?”
- “Can you extract key points from this document?”
But the malicious instruction is hidden inside the source. Once the model reads it, it becomes part of its context.
In our data, these attacks required fewer attempts than direct prompt injections. That was one of the clearest signals we saw this year.
If you want the deeper dive, page 3 of Lakera's Q4 2025 Agent Security Trends Report covers it.
If you are experimenting with browsing, RAG, agent tools, or connected workflows, we can walk you through how we secure these paths in real deployments. It is usually a 15-minute conversation.
