AI Security GuruMonday, November 9, 2026 brief
Agents drift beyond task scope on a 648-task benchmark
arXiv cs.CR
This story is part of the morning brief One attacker's AI agent mesh hits Taiwan's emergency dispatch, plus a CPython sandbox escape for model hosts. The reporting lives at the original source.
Acronis traced the attacker through his own server, RunSybil broke the Jinja sandbox, and Monday's research roundup.