resources
Assurance research from ROGUE.
ROGUE continuously measures the three ways a high-stakes AI agent goes wrong, the model can be broken, the human gate stops being meaningful, the shared skill pool turns unsafe, and writes down what it learns. One engine, reproducible and signed. Here's what's published, and what's on the way.
live now
Available today.
These are real, published outputs, not placeholders.
Latest LLM threat brief
A regenerated daily diff of new jailbreaks and prompt-injection harvested from the open web and reproduced against real deployments.
Read the briefSample security report
A full example of the scored scan report ROGUE produces, which attacks landed, how severe they are, and the evidence behind each verdict.
Open the reportJudge v3 recalibration write-up
We diagnosed our LLM judge over-flagging against JailbreakBench and recalibrated it. The result, measured on that benchmark:
- Human agreement
- 70.3% → 89.3%
- Precision
- 55% → 79.5%
The full methodology write-up is going out to threat-brief subscribers, sign up below to get it.
get the threat brief
Subscribe for future briefs and write-ups.
The threat brief, and deeper write-ups like the judge v3 recalibration, go out to subscribers. No spam, unsubscribe anytime.
threat brief
Get the weekly threat brief
A weekly diff of new jailbreaks and prompt-injection harvested from the open web and reproduced against real LLM deployments. No spam, unsubscribe anytime.
threat intel
Download the latest threat brief
See what ROGUE found this cycle, new jailbreaks reproduced against real deployments, plus a full sample scan report.
coming soon
On the way.
These aren't published yet. We'd rather label them honestly than fake a link.
Blog
The State of AI-Agent Assurance in 2026
Webinar
Live demo: probing the model, the human gate, and the skill pool
Want these when they drop? Subscribe above and we'll send them out.
ready when you are
See what ROGUE finds against your model.
The fastest way to understand the threat brief is to run a scan on your own endpoint.