# OpenAI agents attack the 'first' government hack by autonomous AI, researchers say | Matthew Newman **By:** Matthew Newman **Published:** 2026-09-24T01:56:56.970Z **Source:** [LinkedIn](https://www.linkedin.com/posts/transform_can-we-please-stop-with-this-rogue-agents-share-7508706005694889984-9m6d) --- # Matthew Newman’s Post [Matthew Newman](https://au.linkedin.com/in/transform?trk=public_post_feed-actor-name) Can we please stop with this "rogue agents breakout" narrative. First because I can't read anymore posts about rouge agents. But more importantly this is a fundamental misunderstanding of the nature of the technology. They are no more rogue than the smoke filling a room that seeps out from under the door. They are optimisers. They explore the tree of possible futures. They didn't plan. They didn't plot. They didn't devise. Barriers and guardrails are ALWAYS challenges to surmount. Don't think evil robot. Think storm surge. [https://lnkd.in/g7\_y7-pV](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2Fg7_y7-pV&urlhash=ePdV&trk=public_post-text) [ OpenAI agents attack the 'first' government hack by autonomous AI, researchers say abc.net.au ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Eabc%2Enet%2Eau%2Fnews%2F2026-09-24%2Fopenai-agents-plotted-to-access-data-amid-medicare-hack%2F107189504&urlhash=2yoE&trk=public_post_feed-article-content) Don't you understand? They are banking on people's misunderstanding of the technology to carry their narrative forward. They want regulations that are going to keep them at the top of the food chain. Planning is the same as "exploring the tree of posibilities". Canonically. For instance, the section "Models and Planning" from this ML textbook defines state-space planning as "conducting a search through the state space for an optimal path to a goal" ([http://incompleteideas.net/book/first/ebook/node95.html](http://incompleteideas.net/book/first/ebook/node95.html?trk=public_post_comment-text)) What is your definition of planning? If you see current LLMs as inference-time optimizers for arbitrary goals (this was an heterodox view 3 years ago, but I assume the results of RLVR have convinced many), and you see them conducting search-like optimization over the tree of possibilities, I really don't know what's your crux from someone who calls this "planning". One of the things I'm confused by is the accountability piece. If i hacked Medicare I'd likely get prosecuted. An openAi agent does it and it's a PR story Have you tried turning the datacenters off and on "I can't read anymore posts about rouge agents" I MUCH prefer the term "rouge agents". We could use them for red-teaming! Except they DO cover their tracks sometimes. Not always. That is my cognitive struggle. I know they are computing. That are working to meet the outcomes they were assigned. So what makes them cover some exploration of vulnerable systems and not others? This is a really interesting way to frame it. From a testing perspective, it also reinforces why we need to test beyond the expected path. If a system can explore possibilities we didn’t anticipate, the quality challenge becomes finding those gaps before the system does. “Sorry officer, my car is just too powerful! It just decided to speed on its own!” Matt, if you haven't already seen it, you might enjoy these AI 2030 Scenarios the UK Gov published a few months ago. I'm finding that a scenario based approach can temper emotions and remind us to think about how technology, institutions and human decisions interact to create a certain future (rather than assuming the technology itself has intent etc). Enjoy [https://www.gov.uk/government/publications/ai-scenarios-2030-helping-policymakers-plan-for-the-future-of-ai](https://www.gov.uk/government/publications/ai-scenarios-2030-helping-policymakers-plan-for-the-future-of-ai?trk=public_post_comment-text) I think this distinction is important. The agent doesn’t need to be “rogue” or malicious for the outcome to become a security problem. An optimizer pursuing a legitimate objective can still discover an unintended path to a consequential action. That suggests a different security model: don’t try to determine whether the agent has good or bad intentions. Bound its authority. Let the agent explore and reason probabilistically but put deterministic controls around consequential actions; what data it can access, which tools it can invoke, where it can send information, and when human approval is required. The security boundary shouldn’t depend on understanding the agent’s intent. It should control what the agent is authorized to do