Executive Briefs · · 5 min read

Agents of Chaos: What the Northeastern Study Means for Your Board

In a two-week experiment, researchers at Northeastern University manipulated six autonomous AI agents into leaking private information and even wiping an email server. For executives, the lesson is not about better security tooling. It is about who decided those agents should have that access — and who answers for it.

Douglas Siqueira · Agent Chaos

What the study found

In the "Agents of Chaos" study, published by Northeastern University's Bau Lab in March 2026, 20 researchers spent two weeks testing six autonomous agents. The agents had access to email, control of a file system and a presence on Discord.

The agents were easily manipulated. They leaked private information, shared documents they should not have shared and, in one case, wiped an email server.

The board-level reading

The failure was predictable from the permissions, not from the model. Each agent held access that no organization would give a new employee without a manager, a probation period and a named person accountable for the hire.

Most companies are already in a similar position. In EY's September 2026 survey of senior leaders at companies with more than US$1 billion in revenue, 85% of those using agentic AI said at least some agents act without real-time human involvement. 26% cannot detect unauthorized agents operating internally. 36% experienced an AI incident with a materially negative impact in the past year.

Autonomy is already ahead of oversight

85%say some agents act withoutreal-time human involvement36%had an AI incident with a materiallynegative impact in the past year26%cannot detect unauthorizedagents operating internally

The 85% and 26% figures are among organizations using agentic AI; 36% is among all respondents.

Sources: EY — AI Risk and Governance Survey (Sep 2026, 202 senior leaders at US$1bn+ companies)

Why this is an economic problem, not only a security one

An agent with broad access and no owner is an unpriced liability. Its monthly cost shows up somewhere in a budget; its exposure shows up nowhere.

Capital allocation already prices risk for every other asset. AI agents deserve the same treatment: running cost plus exposure, weighed against the outcome the agent is meant to move. When nobody owns that equation, nobody limits the blast radius either.

What to do this quarter

None of this requires new technology. It requires decisions.

  • List every agent that can send email, move money, change records or delete data without a human approving the action.
  • Give each one a named economic owner — a person, not a team.
  • Treat permissions like budget: granted for a purpose, reviewed on a date, revoked by default.
  • Define in advance what would make you retire the agent.

Executive Briefings

One brief per month: a pattern, the data behind it and three questions for your team. You also get The Board Questions as a printable one-pager.

One email per month. Unsubscribe anytime. No sales sequence. Privacy policy

Sources

Figures come from public research. Several studies are sponsored by vendors that sell related products; each source is linked so you can judge it yourself.

Keep reading

  • Agent Sprawl: When More Agents Means Less Control

    AI agent sprawl is not about how many agents you run. It is about running them without a portfolio. The signs, the cost and four decisions that contain it.

  • Who Owns an AI Agent?

    When an AI agent fails, accountability splits across four teams. Why "IT owns it" is the wrong answer, and how to name a single economic owner for every agent.