AI's Third Wave: Coworkers Break the Security Model That Worked for Agents
Token Security says persistent AI coworkers will hold standing privileges that break human-centric OAuth and audit models.
Token Security CEO Itamar Apelblat argues that persistent AI coworkers break identity controls that still fit session chats and task-scoped agents. He says hosted Anthropic and OpenAI products do not give agents their own credentials, so they often run on human OAuth grants, service accounts, or static bearer tokens, and audit logs record the human. OpenAI product lead Tara Seshan has said useful coworkers need data and infrastructure access and may spawn collaborating sub-agents. Google demonstrated a Workspace agent named Chip with its own account in 2024, but it never shipped.
33