A Microsoft paper shows that editing a handful of dormant neurons makes a model emit an attacker command on a trigger string in recorded agent contexts, without a training run. The authors also report detection results.
Academic and industry research shaping the future of AI security, attack, and defence.
A Microsoft paper shows that editing a handful of dormant neurons makes a model emit an attacker command on a trigger string in recorded agent contexts, without a training run. The authors also report detection results.
Corelight found an MCP server pinned to one release that drifted through 15 dependency sets in five weeks, with the config untouched. 23 of 35 MCP servers it reviewed had a recommended install that resolves dependencies fresh from a public registry.
Darktrace Signal Labs showed that stolen Claude credentials let an attacker steer Claude Code agents through memory stores and pull a 1 GB file out through the Files API. No platform vulnerability is involved, and Darktrace has not seen it in the wild.
An ETH Zurich preprint finds coding agents (three to five, depending on the setting) write less secure code when the repository already contains similar vulnerabilities, and that warnings and security prompts generally reduce the effect without removing it.
Check Point Research (5 Nov 2026) counted 1,608 live ARTEX panels and read 10 exposed databases. Operators mostly asked for progress; one run cost $700 to $1,500 and yielded three usable iOS certificates.