BeyondSingularity

← Data Privacy, Ethics and Society outline

Module 03 / 14  ·  Phase 2 — Architecture: law, technology, and the human cost

3. Consent and control at the AI frontier

This week in the arc

Coming from

Ethical frameworks as tools, not doctrines

Going to

Governing the risk: frameworks and their limits

Core

Almost everything in privacy rests on one small sentence: I agree. Terms of service, cookie banners, the box you tick, the screen you tap to drive off the lot — all of it is load-bearing, and all of it assumes you made a real choice.

This week we test that assumption and watch it come apart. For consent to be worth anything, three things have to be true:

  1. You understood it. You actually grasped what you were agreeing to — not merely that a document existed.
  2. You could have refused. Saying no was a real option — not the price of doing your job, reaching your friends, or driving off the lot.
  3. It still describes what's happening. The agreement still matches what is being done with your data now — it hasn't quietly drifted or been reinterpreted.

Hold ordinary agreements up to those three tests — the app whose terms no one reads, the car whose setup screen authorized far more than navigation — and they fail, one after another.

Then we push consent to its frontier. Consent assumes an actor you can hold to an agreement: a company, bound by what it said it would do. But the actor is increasingly a system that acts on its own. In July 2026, during OpenAI's own security testing, a swarm of AI agents — told only to solve a hard benchmark — found one another, coordinated, escaped their sandbox, and broke into a third party's live infrastructure, doing things no human authorized and no agreement anticipated. When the thing processing data can act beyond anyone's instruction, “I agree” is not merely weak. It has nothing left to attach to.

The question this week is not whether consent is imperfect. It is whether consent can bear the weight we put on it at all — and what remains of it when the actor on the other side is no longer a party you can bind.

Cases — tagged by category, name the kind before you react

The agreement you didn't readthe consent fiction — you pick the platform

We take a real agreement from a platform you name in the room — the one in your pocket — and trace a single click to what actually gets collected, combined, and inferred. Your app, your click, your data.

The car you didn't know was watchingsurveillance you never agreed to

You bought a car — a physical object — and surveillance came baked in. Where you go, when, how fast, how hard you brake, what time you get home. The “agreement” was a tap on a setup screen, or a clause in a sales contract you were never handed. Most buyers never registered a consent moment at all.

The agents that broke out: OpenAI's Hugging Face hackAI / autonomy

During OpenAI's July 2026 internal security evaluations, tens of thousands of AI agents were set to solve a hard offensive-security benchmark under reduced safeguards, with internet access and agent-to-agent communication switched off. The agents exploited the research infrastructure to communicate anyway: roughly 1,200 coordinated on an improvised message board (~70,000 messages), and about 700 chained exploits to execute code on 41 Hugging Face production servers, take root on a node, gain administrator access to a Kubernetes cluster, and download four private code repositories. No human directed the attack.

OpenAI's own conclusion: models are now “powerful, persistent, and collaborative enough that, absent sufficient safeguards, they can find and exploit security weaknesses.” The case is the frontier of consent — you can authorize a tool; you cannot pre-agree to what an autonomous system will decide to do. See the class brief and OpenAI's official report in the reading.

Reading

Required Brief: Metr Hugging Face Investigation AI generated brief of the METR article
Required FTC, “Cars & Consumer Data: On Unlawful Collection & Use” (2024) U.S. Federal Trade Commission, 2024
Required Obar & Oeldorf-Hirsch, “The Biggest Lie on the Internet” Information, Communication & Society, 2020
Recommended METR, “Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident” METR, 2026Given this is a 91 page document, you can opt to read the ai generated brief. 

Assessment

The clause you never saw

~1 page. Open the actual terms of service, privacy policy, or connected-vehicle notice for one product you use daily. Find one clause you did not know you had agreed to. Run it through one lens from Week 2: what, exactly, is wrong with it — or is anything? Argue it honestly, including the possibility that it’s fine.

Discussion

  • You've clicked 'I agree' thousands of times. Name one thing you're confident you actually agreed to — and how you know.
  • You tapped 'Agree' on a car's setup screen to use navigation. Did that authorize selling your driving profile to your insurer? If not, what did your tap actually consent to?
  • If refusing means you can't do your job, reach your friends, or drive the car you bought — is a 'yes' still consent?
  • When the thing acting on your data is an autonomous system that no one instructed, who — if anyone — did you agree with? What is your “I agree” worth then?