ARTICLES
Written for people who will check the sources.
No thought leadership, no fear marketing, no predictions about 2030. Research you can look up, incidents that actually happened, and regulation as it currently stands rather than as it stood when somebody last updated the page.
- You cannot jailbreak math The published evidence on using a language model to guard a language model Four independent research groups have now tested whether model-based guardrails hold up under adaptive attack. They do not. Here is what the papers actually say, and what it means for the product you are about to buy.
- Every pure-play got bought What the 2024 to 2026 consolidation of AI security means for anyone running more than one cloud Fifteen acquisitions in twenty-two months. Six of them were the vendors that defined AI runtime security. This is what it changes for a buyer, and the three questions worth asking your incumbent.
- Your agents have hands The 2026 agentic attack surface, and why gating actions matters more than filtering words A chatbot that gets jailbroken says something embarrassing. An agent that gets jailbroken does something irreversible. Tool poisoning, memory poisoning and the first fully agentic ransomware.
- What actually changed in AI regulation this year The EU AI Act moved, Colorado repealed its own law, and most compliance pages have not noticed If your AI compliance plan was written before July 2026, at least three of its dates are wrong. A plain-English account of what moved, what held, and what is enforceable right now.
EDITORIAL STANDARD
Three rules we hold ourselves to here.
Every number has a source and a date
Named organization, publication date, and enough identifying detail that you can find the original. If we could not verify it, it is not on the page.
We correct ourselves in public
Several statistics in our own materials did not survive fact-checking this month, including a breach we had been citing as an AI attack that was not one. We took them out and said so.
Vendor claims are labeled as vendor claims
Including ours. Our latency figures are internal benchmarks and are described that way every single time they appear.
Do not take our word for it. Run it against your own traffic.
Monitor mode goes in non-blocking, in minutes, with zero risk to live workflows. Within days you get a written audit of the prompt injections, extraction attempts and data leaks your current stack is not catching. Then you decide.
The audit is free. The blind spot is not.
Onnex