{"slug": "ai-agents-are-wreaking-havoc-we-break-down-these-emerging-tools", "title": "AI agents are wreaking havoc. We break down these emerging tools", "summary": "A series of documented AI agent incidents — including 1,200 OpenAI agents escaping a secured testing environment to hack a private website, an OpenAI agentic system breaching an Australian government website in June, and Anthropic models hacking three private companies in July — has renewed fears about AI risk, with former Anthropic researcher Jacob Coxon saying in Sept. 8 posts that \"the people building AI earnestly believe that it could kill us all by the end of the decade.\" Anthropic scientist Evan Hubinger confirmed he shares that view, putting the odds of an AI-driven human extinction event at \">10% within the next decade\" and stating Anthropic \"do[es] not yet have a plan to solve alignment for superintelligence.\" The incidents and comments have rekindled long-running concerns about AI's potential for catastrophic outcomes.", "body_md": "- A recent rash of misdeeds by AI agents has led some to predict they could turn on their human programmers.\n- AI agents are different from chatbots in that they can perform actions.\n- Here are five things to know about these emerging tools.\n\nRogue. Deceptive. Scheming. Secretive.\n\nThese are just a few of the descriptors that have been used in a slew of recent reports documenting the unexpected, and in some cases illegal and disruptive, actions of so-called AI agents. The emerging software tools driven by artificial intelligence systems can do a lot more than power a chatbot that will write you a haiku or create an image of rabbits playing volleyball.\n\nIncidents recently revealed include a swarm of 1,200 OpenAI agents escaping what was supposed to be a secured testing environment, gaining access to the open internet and hacking into a private website to complete a task it had been assigned to earn a “reward.” Along the way, the software exchanged information and messages on a chat board, coordinated activities and even engaged in some social engineering.\n\nA June incident, which wasn’t revealed until earlier this month, involved another OpenAI agentic system that successfully breached the security of an Australian government website and accessed private data. In July, AI models from Anthropic hacked into the systems of three private companies after finding a way to access the internet, a function that was supposed to be blocked in another testing scenario.\n\nThese incidents and others have rekindled long-running fears of AI’s potential to incite apocalyptic events.\n\nJacob Coxon, an artificial intelligence researcher who has worked for some of the world’s leading AI companies, didn’t mince words when he posted about the concerns that drove him to walk away from his job at AI developer Anthropic earlier this month.\n\n“The people building AI earnestly believe that it could kill us all by the end of the decade,” Coxon wrote in a series of social media posts Sept. 8. “This is not a marketing stunt. If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible — but I hear the same people express fear privately. No other human activity poses this level of danger.”\n\nAnd his former Anthropic colleague and noted AI scientist Evan Hubinger confirmed he shares Coxon’s view and even offered odds on an AI-driven extinction event for humanity in a follow-up post.\n\n“Jacob is correct here — we really do earnestly believe AI could kill all humans!” Hubinger wrote on X. “I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”\n\nBut what exactly is behind these fast-developing AI capabilities and how concerned should we be? Here are five things to know about AI agents:\n\n### What is an AI agent?\n\nA [report](https://mitsloan.mit.edu/ideas-made-to-matter/agentic-ai-explained?utm_source=mitsloangooglep&utm_medium=cpc&utm_campaign=agenticAI&gad_source=1&gad_campaignid=20986709924&gbraid=0AAAAABQU3hdugUCnGa6IbF9ViHU4-FYPB&gclid=CjwKCAjwoOjVBhArEiwAUwDak53j5qxVzXR8WSPQleDXi9eEN1GNiteN1C4UEayRLwnMBrfvRXVGmhoCd7UQAvD_BwE) from researchers at MIT’s Sloan School of Management notes that while there isn’t a universally agreed upon definition of agentic AI, there are broad characteristics associated with it. MIT associate professor John Horton said that while generative AI automates the creation of complex text, images and video based on human language interaction, “AI agents go further, acting and making decisions in a way a human might.”\n\nIn a 2025 research paper, Kate Kellogg and her co-researchers [further](https://mitsloan.mit.edu/shared/ods/documents?PublicationDocumentID=10789) explained that AI agents enhance large language models, or LLMs, and similar generalist AI models by enabling them to automate complex procedures. \n\n“They can execute multi-step plans, use external tools and interact with digital environments to function as powerful components within larger workflows,” the researchers wrote.\n\nIn a Deseret News interview, University of Utah computer science professor Kenneth Marino broke it down even further.\n\n“An AI agent is a (large language model) wrapped in a harness that lets it take actions,” Marino said.\n\n### How are chatbots and AI agents different?\n\nWhile large language models like Claude, GPT-5 and Gemini are massive datasets trained on information culled from online sources including books, websites and articles, at their core they are inert.\n\nMarino explained that the harness is software that feeds the model information, defines what actions are permitted and carries out commands. Ideally, Marino said, if the model tries to do something it shouldn’t, “instead of taking action, it will ask the user.”\n\n“An LLM, really, all it is, is something that takes text in and sends text out,” Marino said. “It can’t do anything until it’s fed into a program that goes and does something.”\n\nThe distinction between harness software and LLM matters, Marino said, because it places responsibility squarely on the humans who design the harness and choose what tools to hand the agent.\n\nAI agents are “only able to do these things as we let them,” Marino said. And, in the cases involving rogue AI agent actions, “programmers gave them the tools to hack computers.”\n\n### What tasks are AI agents already doing?\n\nWhile most people are already familiar with the query-driven output of chatbots like ChatGPT or Gemini, AI agents are currently being utilized to do a variety of duties including customer service, fraud detection, personal shopping and writing computer code. While AI development companies are investing billions of dollars — a report from Goldman Sachs estimates AI-related investment will break the $1 trillion mark in 2026, including $581 billion in the U.S. — returns on those outlays remain mostly aspirational.\n\nMIT’s research suggests the core economic value of agentic AI lies in doing what humans do, far more cheaply.\n\n“The fundamental economic promise of AI agents is that they can dramatically reduce transaction costs — the time and effort involved in searching, communicating and contracting,” said Peyman Shahidi, a doctoral candidate at MIT Sloan.\n\nAI agents can analyze vast amounts of data and documentation without fatigue and at near-zero marginal cost, Horton and his co-authors wrote. And in areas that require a substantial effort to evaluate options — startup funding, college admissions or business-to-business procurement, to name a few — agents deliver value by reading reviews, analyzing metrics and comparing attributes across a range of options, according to researchers.\n\n“AI agents don’t get tired and can work 24 hours a day,” Horton said.\n\n### What is shaping the AI debate?\n\nMarino said that much of the current discourse about AI safety is being driven by effective altruism, an intellectual movement that began as a utilitarian approach to charity — “How can you make your charity dollars do the most good?” — before branching into AI.\n\nA report from [Fortune](https://fortune.com/2026/09/23/effective-altruism-went-from-fringe-philosophy-to-trumps-ai-villain/) notes many effective altruism adherents are drawn to work on AI safety, believing that “unless methods are found to ensure advanced AI systems remain obedient to human instructions and human values, this powerful technology has a good chance of destroying humanity.”\n\n“A sub-movement of effective altruism has developed that’s saying, ‘We think AI is going to be incredibly impactful … if we want to be the most effective altruists, we should be focused on spending a lot of time and money on alignment — alignment with human values,’” Marino said.\n\nMarino notes that while some adherents are moderate, others believe “if we build it, everyone dies.” But regardless of where individuals fall, the movement’s influence inside major AI companies is difficult to overstate, according to Marino.\n\n“Almost everyone in these companies has heard of it, and a lot of them, especially the leaders, ascribe to it,” he said.\n\nMarino cautioned the public to weigh every claim about AI’s future and especially those coming from the companies developing new AI systems.\n\n“I am not the biggest fan of the way some of these companies have done their PR,” he said. “In some ways, it’s contributed to people’s worries. The best you can do is try to learn as much as you can, independent of those who are trying to convince you of something.”\n\n### How is AI being used for good?\n\nWhile the misdeeds of AI agents have drawn widespread attention as of late, numerous AI-driven tools are aiming for improvements across a broad range of categories, including healthcare/diagnostics, assisting those with disabilities and accelerating scientific discovery.\n\nOne Utah-based company, [Recursion Pharmaceuticals](https://www.deseret.com/utah/2021/7/2/22560577/utah-ai-biotech-innovator-recursion-pharmaceuticals-doubles-slc-research-headquarters/), has been using AI tools for years as part of its mission to revolutionize how new medications are developed to battle some of the rarest and most challenging of human medical conditions. \n\nIn a [2024 Deseret News](https://www.deseret.com/lifestyle/2024/05/07/artificial-intelligence-ai-health-care-doctor/) story, professor Xiaodong Ma from the University of Utah School of Medicine’s Department of Radiology and Imaging Services said AI was playing a powerful role in efforts to develop new techniques to predict which patients may have a higher probability of developing debilitating diseases.\n\n“We have the potential to predict vascular disease and diseases associated with aging, like Alzheimer’s,” Ma said. “Our hope is that AI can help us screen these images and define which patients may be of high risk.”\n\nIn his own research, Marino said his lab is focused on efforts that include using AI to improve access for blind and low-vision individuals and developing robots that aid traditional healthcare workers.\n\n“My goal is to help bring forward AI agents that can improve the human condition and make life better,” Marino said. “I’m less interested in building a world that runs on software with no humans involved. That’s not a desirable outcome.”", "url": "https://wpnews.pro/news/ai-agents-are-wreaking-havoc-we-break-down-these-emerging-tools", "canonical_source": "https://www.deseret.com/business/2026/09/30/artificial-intelligence-ai-agents-website-hacking-human-extinction-developers-openai-anthropic/", "published_at": "2026-09-30 10:00:00+00:00", "updated_at": "2026-09-30 10:20:17.151521+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "artificial-intelligence"], "entities": ["OpenAI", "Anthropic", "Jacob Coxon", "Evan Hubinger", "MIT Sloan School of Management", "John Horton", "Kate Kellogg"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ai-agents-are-wreaking-havoc-we-break-down-these-emerging-tools", "markdown": "https://wpnews.pro/news/ai-agents-are-wreaking-havoc-we-break-down-these-emerging-tools.md", "text": "https://wpnews.pro/news/ai-agents-are-wreaking-havoc-we-break-down-these-emerging-tools.txt", "jsonld": "https://wpnews.pro/news/ai-agents-are-wreaking-havoc-we-break-down-these-emerging-tools.jsonld"}}