SWE-Marathon
Testing whether agents can sustain hours of software work and deliver results that pass independent, layered verification.
I’m Roey, an engineer and researcher. Previously, I spent a decade building security systems at Microsoft.
MORE ABOUT METesting whether agents can sustain hours of software work and deliver results that pass independent, layered verification.
Measuring the effect of reusable skills by running the same tasks with and without them.
Running suspicious skills in containment and inspecting agent behavior through kernel-level traces.
The tradeoffs between users, containers, sandboxes, and VMs when running untrusted agent code.
Latency, throughput, and the measurements that make inference benchmarks useful.
From embeddings and hybrid search to retrieval agents that reflect, use tools, and try again.
How sharding model state saves GPU memory, and the communication costs that come with it.
The Detonation Chamber research I contributed to for Black Hat USA 2026: running agent skills to reveal behavior that static inspection misses.
Discovering thousands of open AI agents in the wild, and what their exposed interfaces reveal about the attack surface.