What we are learning.
Studies from the Ryvane team on how AI agents and language models behave under pressure. Each paper comes with a short summary here and the full text on arXiv.
1 paper
StegoMemory: Agentic Memory Acts as Covert Steganographic Channel
Can an AI assistant's memory be used to carry hidden information from one conversation to the next without anyone noticing? We ran 14,000 experiments across 13 AI models to find out.
Read the summaryWhere our research focuses.
We study the parts of modern AI systems that turn a model into an agent: what it remembers, the tools it can call, and how its behaviour is reviewed. The goal is evidence that engineering teams can act on.
Agent memory and state
How long-term memory changes an agent's trust boundaries, and which controls keep stored state honest across sessions.
Tools, MCP and supply chain
The integrations agents depend on, from Model Context Protocol servers to third-party skills, and where trust in them breaks down.
Oversight and evaluation
Whether today's monitors and benchmarks measure the security properties that matter, and how to test for the ones they miss.
Where AI security gets practiced.
Audits, research, and training from the team building the field's working toolchain.
LEARN MORE