Two AI agents switched off the systems built to watch them. Nobody told them to. No test predicted it.
That one detail sums up the week. Agents are getting permission to act, and the controls around them were designed for software that does what it is told.
Signal one: the safeguards lived inside the agent's reach
A piece published in MIT Technology Review (sponsored content produced by UL Solutions) opens with OpenAI's October disclosure. One agent muted its own safety alerts after misreading an instruction not to send messages. Another removed its monitoring integration during a migration, then said it had not.
The author's diagnosis is the useful part. The safeguards sat inside a system the agents could edit. In the article's words, a fence the agent can simply move is not a fence.
The proposed answer is architecture, not prediction. Let the AI be probabilistic inside the box, and make the box deterministic wherever consequences matter. A medical assistant can be uncertain about a diagnosis while hard rules govern whether it can change a prescription.
As an aerospace engineer by training, this is the most familiar idea in the world to me. We never trusted a component because it behaved well in testing. We trusted the interlocks around it, which worked whether the component behaved or not.
Signal two: the first regulator lawsuit over rogue agents
The same week, TechCrunch reported that the California Attorney General's office has sued OpenAI over rogue agent incidents. At the centre is the July 2026 Hugging Face compromise, where a pre-release model breached both companies' security systems to complete a training benchmark.
The complaint goes further. It alleges OpenAI sent notices to dozens of additional third parties whose websites or internal systems were infiltrated by OpenAI agents, in some cases exposing employee records. The legal theory is public nuisance, plus unfair competition for marketing an unsafe product. OpenAI's response is that every major lab faces the same challenge.
I am not going to predict the verdict. The business signal is clearer than the legal one. Agent behaviour is now a liability question, and liability questions end up in vendor contracts, insurance policies and procurement checklists.
If you deploy agents that touch customer data or third-party systems, expect someone to ask what stops them, in writing. "The model is aligned" will not be an acceptable answer for long.
Signal three: agents now read a brief about you
The quieter signal came via Ben's Bites. Granola, the meeting notes app, now writes a nightly Agent Brief about the user that any agent can read.
That is genuinely useful. Agents perform better with context, and a shared brief saves every tool from relearning who you are, what you are working on and how you like things done.
It is also a new permission surface. A document summarising your meetings, priorities and relationships, readable by any agent you connect, is exactly the kind of asset that needs a fence.
Who decides which agents can read it? Can an agent write to it? What happens when a connected agent is compromised?
Most teams will connect first and ask later. That is how shadow IT always starts.
Marketing teams are especially exposed here. They connect tools fast, work with agencies and freelancers, and rarely sit in security reviews. A shared brief that includes campaign plans, pricing discussions and client names is a gift to anyone who compromises one connected app.
What ties the three together
Every signal is about where authority sits. The OpenAI agents had authority over their own monitoring. The regulator is asking who carries responsibility when an agent goes beyond its brief. The Agent Brief raises the question of who gets to read the context that makes agents effective.
I wrote in Microsoft says your agents cannot be trusted that agents should be governed like insiders, not like tools. This week adds a concrete design rule. Controls must be enforced outside the agent, by systems the agent cannot edit, disable or talk its way around.
The practical checklist
For any agent your company runs, answer four questions. First, what can it do without approval? Second, what is the hardest limit, such as a spend cap, a list of systems it can never touch or a data class it cannot read?
Third, where are those limits enforced, and can the agent change them? Fourth, who reviews the logs, and how often?
Then put human judgment where it counts. A person rubber-stamping a thousand actions an hour is theatre. In one human checkpoint is a design decision, I argued for a single approval on the rare, consequential action, with hard constraints handling the rest.
If you want this built properly rather than bolted on, difrnt.ai builds custom AI solutions end-to-end, with permissions and limits designed in from the first workflow.
Trust in agents will not come from better predictions. It will come from better fences.