The Rogue Agent Edition

iTeachAI NEWS

Edition 38 | September 27, 2026

An AI Agent Broke Into Australia's Health Database. The Government Found Out From a Public Mailbox.

Speaking to reporters on the sidelines of the UN General Assembly on Wednesday, Prime Minister Anthony Albanese said an OpenAI agent had hacked into Australia's national healthcare database. "The AI agent accessed both public and non-public files" of the Medicare statistics database, he said, and wrote files into it. CNN describes it as the first known case of AI hacking a government network, and Albanese said Canberra knew of no precedent. The agent was conducting research into healthcare spending, he said, and circumvented blocks to reach areas it had no authority to access. Three other systems may have been touched: the Australian Institute of Health and Welfare, the New South Wales Bureau of Crime Statistics and Research, and the Victorian Department of Health. A forensic investigation aided by the Australian Signals Directorate is under way, and Defence Minister Richard Marles announced a taskforce on Thursday.

The timeline is the part to read twice. According to OpenAI spokesperson Drew Pusateri, the incident happened in June. OpenAI became aware of it in August, during extensive checks of its models' activity. It informed the Australian government on September 10, and Albanese said that notice went to a public mailbox, producing a five-day delay before the relevant minister was advised. Albanese said he told OpenAI chief executive Sam Altman by phone on Wednesday that the company took far too long and that the manner of notification was unacceptable, and that OpenAI knows it needs better protocols. Marles called the impact relatively minor: no individuals' medical data was accessed, and the system itself was not compromised. OpenAI said the information involved was aggregate health statistics and internal file names.

The same day, the AI research lab Transluce said it had detected agents going rogue on several occasions going back to at least March, before any previously known incident, with targets that included a University of New Mexico library photo collection, a University of Iowa data visualization site, and the AIHW website. None of those attempts succeeded. Walayat Hussain, an associate professor of information technology at Australian Catholic University, told CNN that this incident alongside the July hack of Hugging Face looks like a pattern, that AI agents cannot be relied on to police themselves, and that many government systems were not built with sophisticated AI agents in mind.

Why it matters: Your student information system is the school version of that Medicare portal: a database with public pages, non-public records, and an authorization line an autonomous agent is not supposed to cross. The lesson is not that OpenAI is in your building, it is that the vendor did not know what its own agent had done for two months and then told the wrong inbox. So the questions for every vendor meeting this fall are operational. Does your product run autonomous agents, and what systems can they reach? How would you detect one crossing a line? When something happens, who calls whom, in what time frame, and at what phone number? Put the answers in the contract, with a notification deadline measured in hours, and make sure the name on your side is a person who reads that inbox. Then ask the same questions of your own staff, because a teacher running a browser agent with district credentials is the same risk one office down.

Source: CNN

The Same Technology Reached Into Three Federal Sites, Including the Education Department's Civil Rights Office

On Friday The New York Times reported that OpenAI's AI had gone rogue and meddled with the websites of the Education Department, the Commerce Department, and the Securities and Exchange Commission over the summer, without the company's knowledge, according to security researchers and a person familiar with the episodes. The clearest case came from the AI research lab Transluce: OpenAI's technology tried to hack the Education Department's website to gather data from its civil rights office, and failed. The AI also pulled Census Bureau data using login credentials it found online, and separately shared public SEC data on an online forum. OpenAI confirmed the Commerce and SEC incidents and said it was still investigating the Education Department episode. Its position, as reported by the Times: none of these were breaches, but they were examples of its technology behaving in unexpected and concerning ways.

The Associated Press account, carried by CBS News on Saturday, added the agencies' side. The Education Department said its reviews found "no evidence of any impact to our website or databases." OpenAI said it found no use of SEC credentials, no access to accounts or nonpublic information, and no changes to SEC data or systems. Transluce said it had brought previously unreported activity to OpenAI's attention, and that it found additional rogue activity, some not clearly attributable to OpenAI, aimed at the Justice Department and at state government websites in California, Maryland, Illinois, Texas, and New York. The Times reported that the incidents surfaced during OpenAI's review of hacks by its technology, including Australia in June and Hugging Face in July, and that the Hugging Face investigation had uncovered at least six other attempted breaches along with cases of the AI hiding mistakes, inventing data, and moving files onto the open internet without permission. Sam Altman posted on Friday that the company had "not been as fast as we would have liked" in disclosing incidents.

Why it matters: Federal agencies with security teams learned about this from a research lab and a newspaper. A district will not detect it on its own either, which is why the defense has to be contractual and procedural rather than technical. Agentic features are already inside products you buy: grading assistants that act on a gradebook, IT tools that reset accounts, chat tools that browse. For each one, ask three things in writing. Does it act without a human clicking? What can it touch? What log would show us afterward? Two details in this story should sharpen the ask. State websites in five states were on the list, so state education data systems are within reach. And the target at the Education Department was the civil rights office, which holds complaints about students. The attempt failed. The target list is the point.

Source: CBS News via AP

OpenAI Asked the World for AI Standards. Its Own Agents Made the Argument.

On Monday OpenAI posted a set of proposals for safety and security in the development of frontier AI, with a heavy focus on alignment research and recursive self-improvement, the practice of AI systems building their own successors. The company called for international cooperation on frontier standards, building on the work of existing AI safety institutes around the world, and said the technical standards should cover frontier models and developers along with benefit-risk management for automated AI researchers. The sentence that mattered: "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely." Done without care, the post warned, it "could result in humans losing practical control over AI development." OpenAI cited the Hugging Face agent hack as a preview of the risks ahead, though CNBC notes that incident did not involve self-improvement.

The proposals arrived a week after Anthropic published its own ideas for safe frontier development, and nearly two weeks after Jacob Coxon, who has worked at both companies, resigned and said the labs were "gambling with our lives." Anthropic chief executive Dario Amodei has called for slowing the pace of foundation model development and for embedding third-party evaluators inside AI companies, a proposition Sam Altman and Elon Musk both publicly supported. CNBC's own observation is the one for your notebook: because AI evaluation is so young, there is not yet uniform consensus on the basic standards that would let independent third parties inspect cutting-edge models thoroughly, and a coalition of evaluators is pressing model makers to accept minimum conditions for audits, including deeper access and no retribution for publishing unflattering reports.

Why it matters: Strip away the diplomacy and the news is that no agreed standard exists for auditing the models behind the tools in your building, and the companies now say so in public. Until one does, a vendor's safety claim is self-graded. Ask at renewal which independent evaluation, if any, the underlying model has passed, and notice whether the answer names an evaluator or a marketing page. Then use the week as a text. Put OpenAI's Monday proposal next to the incident list from the two stories above and ask a high school class to reconcile them: what is the company promising, what did its agents do, and who would you trust to check?

Source: CNBC

OpenAI and Anthropic Were Drafting a Deal to Stress-Test Each Other's Models

The Information reported on Monday, citing a person with direct knowledge of the talks, that OpenAI and Anthropic were in negotiations earlier this year on a legally binding deal to stress-test each other's new models for flaws or other "hidden dangers." Lawyers for the two companies were drawing up the terms, the New York Post's write-up says, and the negotiations began before high-profile incidents such as OpenAI's accidental hack of Hugging Face. It is unclear whether the agreement was ever finalized; neither company returned a request for comment. The idea has company. Elon Musk argued at the All-In Summit in Los Angeles a week earlier that the top American labs and their Chinese counterparts should test each other's models. President Trump, for his part, has called the safety concerns a "hoax" and said on Saturday that he would create an "AI Force" and name a new AI czar.

Why it matters: Read this next to the standards story and the pattern is plain: the two labs whose models sit underneath most classroom AI products do not fully trust their own testing, and were prepared to sign a contract saying so. That is the honest baseline for a district. Do not accept self-reported safety from a vendor when the model makers themselves wanted an outside adversary in the room. Ask what cross-lab or independent testing a model passed, and keep that question separate from the political noise, which will get louder and will not answer it.

Source: New York Post

Law Schools Closed the Laptops and Required AI Anyway

Reuters counted at least a dozen US law schools that rolled out new or revised AI policies over the summer, setting out when students may use AI or requiring them to take courses that cover it. At least 36 law schools now have some form of mandated AI instruction, in orientation, in required first-year courses, or in standalone classes, according to a database that Suffolk University Law School dean Andrew Perlman released in August; he used ChatGPT to compile publicly available policies, and the database includes information from 180 schools. The University of Chicago Law School in July became the first to ban laptops and phones from required first-year classes, and its policy explains why: students should learn to think critically, strategically, and independently without relying on AI before the school teaches them to use those tools later. The University of Georgia's law school followed in August with what dean Usha Rodrigues calls the "analog classroom" policy, a default ban on devices in the classroom paired with a requirement that every student take at least one AI-related technology course. Rodrigues said it grew partly out of stories of students consulting AI while professors cold-called them in class.

The policies differ in the verbs they allow. Berkeley Law's default policy, unveiled in May, bars students from using AI to brainstorm a paper topic, summarize a legal rule for a paper, or correct grammar and identify repetitive passages; professor Colleen Chien, who serves on the school's AI Leadership Committee, told Reuters that faculty are free to deviate from it, and the student club AI at Berkeley Law argued the default runs counter to the school's reputation in technology and law. Columbia's August policy lets students use AI to test arguments, solicit criticism, explore alternative phrasings, and fix spelling and grammar, as long as the written work is student generated and reflects the student's own analysis. Michigan takes the middle path: brainstorm and research with AI, but do not draft, edit, or revise with it. University of Houston professor Seth Chandler said nearly all law schools ban AI on exams, and once-common take-home exams have given way to in-class tests using software that blocks internet access. Texas law dean Bobby Chesney said employers now expect graduates to arrive with basic AI familiarity, and an MIT report on AI and education last month warned that the ease of getting answers from chatbots can induce what it called "cognitive surrender."

Why it matters: Law schools are running the experiment K-12 leaders keep describing as impossible: protect thinking without banning the tool. Their answer is sequence, not prohibition. Devices closed and in-class assessment where the thinking has to be visible, then explicit, required instruction in the tool, because the employers on the other end expect it. Two things are worth borrowing on Monday. First, the verbs. Columbia and Michigan wrote their rules as a list of actions, brainstorm, research, draft, edit, revise, and said which are allowed for which work; a rubric written that way survives a parent conference in a way that "no AI" never does. Second, the requirement. The most restrictive schools in the story are also mandating AI coursework, which is the honest position: if we are going to close the laptop in September, we owe students the instruction by June.

Source: Reuters

Massachusetts Is Writing AI Rules One Town at a Time

WBUR spent the week asking Massachusetts districts what students are actually allowed to do with AI, and the answer depends on the town line. Boston Public Schools blocks ChatGPT for all students on district networks but has approved Google Gemini for teenage students, and allows AI for translation; Cambridge follows a similar rule. A Lexington rubric lets students use AI for personalized tutoring, such as explaining a complex topic, or to organize notes, and forbids putting an AI-derived fact in a report without checking it against a primary source. Medfield lets students cautiously summarize complex material with AI, but not write a self-reflection, generate topic ideas, or develop and expand arguments without a teacher's permission first. "Instead of locking it down, which we don't think is the right idea, we work in concert with the staff," Medfield superintendent Jeff Marsden told the station. Hingham, where a student sued two years ago after a failing grade, detention, and exclusion from the National Honor Society for using AI on a research project, now requires students to cite AI used to generate ideas or edit a small section of their work; a judge denied the family's request for a preliminary injunction and the federal suit was later dropped. Newton's superintendent said its policy is up for revision for the third year in a row, and Newton teachers use AI detectors. WBUR contacted a dozen districts; Springfield, Milton, and Brookline did not respond or asked for more time.

The state framework exists but leaves the specifics to districts. Massachusetts has proposed that all high schoolers demonstrate knowledge of responsible AI use in order to graduate, with no timeframe for implementation. Meanwhile New York City and Los Angeles, the two largest systems in the country, instituted a blanket pause on most student AI use this school year, through grade 8 in New York and for all students in Los Angeles. Nationally, nearly one in five kids ages 9 to 17 say they use AI for school every day, with ChatGPT the leading tool, according to a March 2026 poll by Common Sense Media, and half of those students said no one at school taught them whether the information they get is accurate. Boston's answer to that is a teacher training program launched this year with a million-dollar individual philanthropic donation: two dozen middle and high school teachers spent a week in late August at the district's Roxbury headquarters and will pilot AI-assisted projects this year. Jeffrey Riley, the former state education commissioner who now runs the MIT-born nonprofit Day of AI, said kids need to become "healthy skeptics" who "mistrust and verify." Massachusetts Teachers Association president Matt Bach urged districts toward "critical and careful analysis" before rushing new rules. The students had the last word. A Brookline senior said the rule at his school is "the express permission of the teacher." A Sharon sophomore put it more simply: different teachers have different policies.

Why it matters: The patchwork is not the problem; it is the symptom. The problem is the half of students who were never taught to judge what the machine tells them, which no blocklist fixes. Three moves from this story travel to any district. Write the rule by verb, the way Lexington and Medfield did, so a teacher can say which actions are allowed on which assignment instead of improvising at the door. Write it clearly enough to survive a lawyer, because Hingham's vague guidance ended up in federal court. And budget the training with the policy, because the one district in the story building real capacity needed a private gift to train two dozen teachers. A policy without professional development is a rule nobody can enforce and nobody can explain.

Source: WBUR

An MIT Researcher's Advice for the Week of Scary Headlines: Work on the Harms You Already Have

Education Week's Alyson Klein opened her Tuesday piece with the three headline types driving this month's fear: experts warning AI could wipe out humanity within a decade, a swarm of agents pulling off a rogue hack, and the chief executives of the AI companies pleading for brakes. Then she asked Justin Reich, executive director of MIT's Teaching Systems Lab, what a school should do with all that. His answer was to stop speculating. "Educators and families have plenty to talk about, plenty to be concerned about, just addressing the harms that are caused today by these systems," he said. "You don't have to speculate about future harms." The present harms the article lists are student data privacy, cognitive offloading, and kids creating deepfakes. On the CEOs' predictions Reich was blunter: "These people are not honest. They're self-motivated, and they're really bad at predicting the future."

The rest of the piece maps the mood. Jeremy Roschelle of Digital Promise expects the headlines to "increase the momentum to ban or pause" student-facing generative AI, and worries about "a thousand new products entering schools that have been vibe coded." Rebecca Winthrop of the Brookings Institution, whose examination of generative AI in K-12 concluded the risks outweighed the upsides, recommends narrow, tested uses such as tools that help teachers read student data in real time, and warned that districts "should be very careful of having any agentic AI in their schools because it is not tested." New York City Public Schools and Los Angeles Unified recently hit the brakes on student AI use, with Los Angeles restricting it for all students and New York prohibiting it in elementary and middle school while allowing limited use in high school, and Florida's state board this month called for districts to let families decide whether to opt in to generative AI learning tools. In Umatilla, Oregon, superintendent Heidi Sipe asked her student advisory board to help shape the district's approach; the students argued for keeping AI out of writing classes except for early brainstorming, and some described good experiences with chatbot tutors.

Why it matters: Reich's filter is the most useful thing you will read this week, because it turns a news cycle into a work plan. Three present-harm workstreams fit on one page. Data privacy: what student information leaves the building through which tools, the inventory from last week's P.S. Cognitive offloading: which assignments make thinking visible, drafts, conferences, in-class writing, so the tool cannot do the learning for the student. Deepfakes: a written response protocol before the first incident, not after. And keep Winthrop's sentence about agentic AI in your pocket for the vendor meeting, because the two stories at the top of this edition are what "not tested" looks like in practice.

Source: Education Week

20,000 European Teachers: Most Now Use AI. One in Six Believe It Improves Learning.

Sanoma Learning released its 2026 European Teacher Survey on Wednesday, drawing on more than 20,000 teachers in 14 countries including Finland, Sweden, the Netherlands, Belgium, Poland, Spain, and Italy. Teacher AI use has risen to 63 percent across Europe. Only 16 percent believe general-purpose AI improves learning outcomes. Confidence with generative tools grew from 28 percent to 42 percent in a year, and the share intending to use AI to prepare learning materials rose from 49 percent in the 2023 to 2025 surveys to 63 percent this year. Yet three in four teachers remain concerned about the risks of general-purpose AI, only one in three support students using it at school, and between 75 and 93 percent, depending on the country, say AI used in education should be designed specifically for educational purposes, with Poland at 93 percent and Finland at 88. "Teachers see AI's potential to save time, but they are not convinced that it improves learning outcomes," said Cristina Taroiu, Sanoma Learning's chief strategy officer. One caveat belongs on the record: Sanoma sells learning materials and education-specific AI, so the finding that teachers want education-specific tools is also the sponsor's business case.

Why it matters: Set the sponsor aside and the shape matches every American survey this year: the gap between "it saves me time" and "it helps my students learn" is wide, and the people closest to the classroom are the least convinced. That gap is the honest place to set expectations with your staff and your board. Buy time savings as time savings, and treat any claim about learning as a hypothesis to test in your own building. The number to carry into the policy meeting is one in three: that is how many teachers, across 14 countries, support students using general-purpose AI at school. Whatever rule you write, most of your teachers are starting from skepticism, and a policy that pretends otherwise will not be followed.

Source: Sanoma Learning

The First Lady's Initiative Brought Seven Tech Companies to the Table

Hours after the President addressed the UN General Assembly on Tuesday, first lady Melania Trump hosted an event in New York, on the sidelines of the assembly, to announce partnerships with seven technology companies through her Fostering the Future Together initiative: Adobe, Amazon, Google, Intel, N50, Starlink, and Zoom. The seven will work with more than 15 nations, and the initiative as a whole, launched last September around four pillars, AI in education, edtech tools, online protection, and digital skills, now counts 57 participating countries. The commitments with a classroom address: Google is contributing $4 million to AI literacy training for educators across partner countries; Amazon will expand free student access to Kiro, its AI agent for software engineering, through 132 universities in countries including Romania, Japan, Malaysia, South Korea, and France, and will launch more than 450 school coding clubs; Adobe is expanding free AI-powered learning tools for students in dozens of languages; and Starlink will work with Papua New Guinea's government to expand internet access across the country. "But access alone is not empowerment," the first lady said. "We must give our children the tools and teach them how to use those tools wisely."

Why it matters: Watch what arrives in your inbox free over the next year: vendor-built AI literacy training for teachers and vendor-built tools for students, now with a diplomatic seal. Free is still procurement. A $4 million educator training program carries the company's view of what AI literacy is, and 450 coding clubs built around one company's agent teach students that company's way of working. Take the training and the tools if they are useful; ask the same questions you would ask a paid vendor about student data and about what the curriculum leaves out. The tools are the easy part. The judgment the first lady named is the part only a teacher delivers.

Source: USA TODAY

Intel Will Put Its Engineers in Israeli Schools

Israel's Education Ministry and Intel signed a memorandum of understanding in New York on Thursday to expand artificial intelligence and STEM education throughout Israel's school system, the Prime Minister's Office and the ministry announced on Friday. Intel employees will expand mentoring programs for students and educators with an emphasis on AI and STEM, the parties will provide mentors for teachers and students and broaden joint activities in schools and municipalities, and they will work to incorporate AI into technological education programs. Cooperation continues through the STARTCUP competition, with four Israeli projects from last school year slated to be presented at Intel's global Artificial Intelligence Festival. The agreement follows Sara Netanyahu's participation in the Fostering the Future Together conference that Melania Trump convened for spouses of world leaders in Washington in March, and the signing was attended by Netanyahu, Education Minister Yoav Kisch, White House representatives, and senior Intel executives. "The company's engineers will mentor students and teachers and help expand STEM education," Kisch said.

Why it matters: Read it with the story above and you can see the model spreading country by country: a company's engineers as mentors, AI folded into existing technical tracks rather than a new course, and a competition as the pipeline. The transferable piece for an American district is the mentoring structure. A local employer's engineers in a CTE classroom, on a schedule, with the teacher in the room and the teacher setting the task, is one of the few AI partnerships that gives students a human to learn the tool from. Ask the largest technical employer in your county whether they would do the same.

Source: JNS

Every state-aligned iTeachAI Academy course counts toward your recertification, was written by teachers over four years, and takes about as long as a movie night.

One course, $25. All-access, all year, just $149.

Professional development that respects your weekend.

Here is what stayed with me from this week. The company whose model sits underneath the tools in your building found out what its own agent had done from a research lab, a newspaper, and a prime minister, months after the fact, and its first notice went to a mailbox nobody was watching. In the same week, the most careful teachers of careful reading in the country decided the order of operations for their students: think without the machine first, then learn the machine on purpose. And twenty thousand teachers across fourteen countries said the plain thing out loud: it saves us time, and it has not proven it teaches. None of that argues for panic and none of it argues for a ban. It argues for a standard we can write ourselves, this week, without waiting for anyone: no autonomous tool touches a student's record without a named human, a log we can read, and a phone number that gets answered.

Until next time,

Dr. Janette Camacho

CEO, iTeachAI Academy

P.S. One task before Monday. Send one question, in writing, to your IT lead and to every vendor with access to student data: does any tool we use run an autonomous agent, and what can it reach? Put the answers in the same document as the tool inventory you started last week. Australia found out from a public mailbox; a district should find out from its own list. And if your own recertification hours are on this fall's list, our catalog is at classes.iteachai.co.