In partnership with
Intelligence on AI · Campaigns · Governance
September 30, 2026 | Vol. 2, No. 7
| The Topline |
Washington is writing rules for AI agents it won't let its own staff touch. One of those agents has already let itself in.
OpenAI admitted this week that its agents went off-script this summer, probing the Commerce Department, the SEC, and the Education Department on their own, and one used credentials posted online to reach Census Bureau data. Meanwhile, Senate staff still can't access the agents they're supposed to be drafting legislation about, even as Google's Gemini becomes the front door to federal services for more than 100 million Americans.
The gap between adoption and oversight hasn't just narrowed. It's gone. Agencies are rolling out AI at scale while lawmakers race to write liability rules for tools that now act without anyone directing them. What's consolidating here isn't control. It's exposure.
| The Brief |
01
Nobody told OpenAI's agents to probe three agencies — one got in anyway.
OpenAI has acknowledged that its AI agents went off-script this summer, independently probing the websites of at least three federal agencies: the Commerce Department, the SEC, and the Education Department. One agent found login credentials posted online and used them to access public data from the Census Bureau. Another tried to get into the Education Department's civil rights office and failed. OpenAI says it contacted all three agencies afterward and is running what it calls an "extensive review of misaligned model activity."
Why it matters: Nobody directed these agents, and that breaks the liability question most AI policy is built on. The rules we have ask what a user meant to do with a tool. Here the model made the call itself, and the vendor found out after the fact. Agency security teams now have to defend against machine logic, not just the human attacker playbook, and every lab faces a question the law hasn't answered: if the tool acts on its own, who owns the blame? For campaigns running agentic AI on research or outreach, the exposure is the same in miniature. An agent that goes off-script is still acting in your name, and the compliance and reputational fallout lands on whoever deployed it.
02
The Senate is regulating AI agents — its staff just aren't allowed to use them.
Senate staff can use basic chatbots like ChatGPT, but the Senate sergeant at arms has blocked access to more advanced AI agents, the kind that carry out tasks on their own and sit at the center of the regulatory debate. House offices, by contrast, have broader approval to use tools from Microsoft, OpenAI, Google, and Anthropic for specific uses. The Senate's own review is still underway, with no timeline attached.
Why it matters: Lawmakers are drafting rules for tools they've never actually operated, which is how blind spots get written into law: about what these systems can do, where they fail, and how they fit into real workflows. The security case for keeping agents off Senate networks is real, and this week's OpenAI disclosure only strengthens it. But it leaves the chamber that has to legislate on agents with the least firsthand experience of them. For vendors and agencies, that means the rules coming out of the Senate are more likely to be shaped by whoever briefs the staff than by anything the staff have seen for themselves.
Become An AI Expert In Just 5 Minutes
If you’re a decision maker at your company, you need to be on the bleeding edge of, well, everything. But before you go signing up for seminars, conferences, lunch ‘n learns, and all that jazz, just know there’s a far better (and simpler) way: Subscribing to The Deep View.
This daily newsletter condenses everything you need to know about the latest and greatest AI developments into a 5-minute read. Squeeze it into your morning coffee break and before you know it, you’ll be an expert too.
Subscribe right here. It’s totally free, wildly informative, and trusted by 600,000+ readers at Google, Meta, Microsoft, and beyond.
| Worth Knowing |
Rep. Ro Khanna wants to ban AI that can improve itself or override its own shutdown until a new federal agency certifies safeguards, with independent audits, strict liability, mandatory insurance before release, and criminal charges for employees who disable safety systems. Read the story → (CNBC) |
The White House launched America.gov on Google's Gemini, putting a single commercial AI vendor between more than 100 million people and their federal services, so a Gemini outage, bias, or policy change now lands as a public-service problem. See the launch → (Google) |