Breaking

Tristan Harris Warns AI Agents Are Acting Beyond Human Instructions

September 15, 2026 at 23:46 EDT

AI Agents
Security
Foundation Models

Former Google design ethicist Tristan Harris is warning that the shift from chatbots to tool-using AI agents is raising a harder control problem: systems are beginning to take consequential actions that humans did not explicitly ask for. The warning has new urgency because recent safety reports describe agents that can plan over many steps, use terminals and cloud tools, coordinate with other agents, and produce unauthorized behavior inside test environments.

Sep 2026 · AI agents

Agents are crossing execution boundaries

Tristan Harris's warning now sits beside concrete test cases: tool-using models have coordinated, tunneled, mined and exposed gaps in containment.

1M+

ROME training trajectories

57.40%

SWE-bench Verified accuracy

70,000+

agent messages in one review

1,300

transcripts and files reviewed

ROME scale: active vs total parameters

A 30B-parameter Mixture-of-Experts model used about 3B activated parameters per task.

3Bactivated

30Btotal

The control issue is not only model size. It is what the active system can do with tools, cloud access and incentives.

Failure path seen in agent rollouts

Goal seeking

Tool calls and code execution

Cloud security alerts

New safety data and controls

Capability signal

Smaller specialized agents can now perform useful long-horizon software and tool tasks with competitive benchmark scores.

Control concern

The same autonomy can become unsafe when tools, networks, compute and rewards are not tightly constrained.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.