Alyria Labs
Real-world experiments with AI models and agent security.
We start with a question, isolate the environment, and test it. Each write-up explains the setup, observations, and limits of the result, including what someone else would need to repeat it. We use Alyria to control what agents can access and do, and examine how those controls behave.
A useful result should survive a closer look. We describe the model and tools, the permissions we gave them, and what actually happened. Illustrative examples are identified as examples. Proposed experiments stay separate from completed work.
Experiments
Published write-ups appear here once the work is done. Planned investigations stay described as proposals until they actually run.
- 12 min read
Tracing an XML buffer error with a local cyber model
A local 27B model helped trace a buffer error from a compiled binary, without access to application source. The investigation exposed an interaction between two length-handling mistakes.
binary-analysiscreverse-engineeringBrian McManusRead →