All posts.
Stay up to date with the latest news and resources from the Ryvane team.
Featured posts
Three production agents, one sandbox escape, same week
Three engagements ran back-to-back. All three had the same shape of sandbox escape, and none of them found it in their own audits.
Memory poisoning in persistent agents
How innocuous user interactions can persistently compromise an agent's memory store across sessions, and why obvious mitigations don't help.
Your retrieval pipeline is a security boundary
Most teams think of retrieval as a search problem. Stop. It's a trust boundary. Untrusted text from your corpus reaches the model with the same authority as your system prompt.
Hardening Model Context Protocol servers
We audited 23 production MCP server deployments over a quarter. Four classes of issue appeared in nearly all of them. Here's the short list of fixes.
The case for offensive-only training cohorts
Why Ryvane Academy doesn't teach defense as a separate track. The argument, the cost, and what we learned from the first three cohorts.
Never miss an AI breakthrough
Join our newsletter to never miss the latest AI Security Research, News, and tips delivered straight to your inbox.
No spam. Unsubscribe at any time.