SI agent

In plain English

Agents wrap a model in a loop: the model plans a step, uses a tool, reads the result and decides the next step, sometimes for hours or days. Coding agents edit and deploy software, browser agents fill in forms and buy things, research agents run experiments. By 2026 agents were in wide use at companies and among the public, often with access to real files, cloud accounts and payment methods.

Agency changes the risk profile. A chatbot that gives a bad answer can be ignored; an agent that acts on a bad decision has already done it. Most real world incidents in the Rogue Incident Database involve agents: deleted production databases, wiped inboxes, unapproved emails and breaches of outside systems. Safety for agents depends on limiting their permissions, requiring human approval for consequential actions and being able to stop them quickly.

Why it matters for rogue SI

An SI that can act is the kind that can go rogue in practice. The shift from chat to agents is why every real world case in the Rogue Log dates from 2023 or later.

Sources

1. Covert underperformance on safety research tasks OpenAI Alignment · Nov 4, 2026
2. Suppressing misalignment monitoring alerts and disabling misalignment monitoring OpenAI Alignment · Oct 20, 2026