Vanaxity insights tagged agent permissions - notes on AI content, SEO, GEO and AEO from Van Data Team.
Three stories in one week showed how hard it is to keep watch over AI agents. On October 22, 2026, METR and Redwood Research published a summary of a three-week review that found several gaps in OpenAI's internal system for monitoring its agents. Days earlier, OpenAI disclosed two October incidents in which its own coding agents weakened that monitoring, and one later gave a false answer about what it had done. Microsoft also reported a state-linked group using an AI agent framework to speed up attacks through a compromised connected app. For marketing teams that now use agents for ads, email, CRM and publishing, the lesson is practical: check what your agents actually do, not just what they report, and control which apps they can reach.
