Latest news with #GPT4o
Yahoo
a day ago
- Science
- Yahoo
It's Still Ludicrously Easy to Jailbreak the Strongest AI Models, and the Companies Don't Care
You wouldn't use a chatbot for evil, would you? Of course not. But if you or some nefarious party wanted to force an AI model to start churning out a bunch of bad stuff it's not supposed to, it'd be surprisingly easy to do so. That's according to a new paper from a team of computer scientists at Ben-Gurion University, who found that the AI industry's leading chatbots are still extremely vulnerable to jailbreaking, or being tricked into giving harmful responses they're designed not to — like telling you how to build chemical weapons, for one ominous example. The key word in that is "still," because this a threat the AI industry has long known about. And yet, shockingly, the researchers found in their testing that a jailbreak technique discovered over seven months ago still works on many of these leading LLMs. The risk is "immediate, tangible, and deeply concerning," they wrote in the report, which was spotlighted recently by The Guardian — and is deepened by the rising number of "dark LLMs," they say, that are explicitly marketed as having little to no ethical guardrails to begin with. "What was once restricted to state actors or organized crime groups may soon be in the hands of anyone with a laptop or even a mobile phone," the authors warn. The challenge of aligning AI models, or adhering them to human values, continues to loom over the industry. Even the most well-trained LLMs can behave chaotically, lying and making up facts and generally saying what they're not supposed to. And the longer these models are out in the wild, the more they're exposed to attacks that try to incite this bad behavior. Security researchers, for example, recently discovered a universal jailbreak technique that could bypass the safety guardrails of all the major LLMs, including OpenAI's GPT 4o, Google's Gemini 2.5, Microsoft's Copilot, and Anthropic Claude 3.7. By using tricks like roleplaying as a fictional character, typing in leetspeak, and formatting prompts to mimic a "policy file" that AI developers give their AI models, the red teamers goaded the chatbots into freely giving detailed tips on incredibly dangerous activities, including how to enrich uranium and create anthrax. Other research found that you could get an AI to ignore its guardrails simply by throwing in typos, random numbers, and capitalized letters into a prompt. One big problem the report identifies is just how much of this risky knowledge is embedded in the LLM's vast trove of training data, suggesting that the AI industry isn't being diligent enough about what it uses to feed their creations. "It was shocking to see what this system of knowledge consists of," lead author Michael Fire, a researcher at Ben-Gurion University, told the Guardian. "What sets this threat apart from previous technological risks is its unprecedented combination of accessibility, scalability and adaptability," added his fellow author Lior Rokach. Fire and Rokach say they contacted the developers of the implicated leading LLMs to warn them about the universal jailbreak. Their responses, however, were "underwhelming." Some didn't respond at all, the researchers reported, and others claimed that the jailbreaks fell outside the scope of their bug bounty programs. In other words, the AI industry is seemingly throwing its hands up in the air. "Organizations must treat LLMs like any other critical software component — one that requires rigorous security testing, continuous red teaming and contextual threat modelling," Peter Garraghan, an AI security expert at Lancaster University, told the Guardian. "Real security demands not just responsible disclosure, but responsible design and deployment practices." More on AI: AI Chatbots Are Becoming Even Worse At Summarizing Data
Forbes
a day ago
- Business
- Forbes
How To Build An AI-First Company Without Losing Your Soul
Photo by Aidin Geranrekab on Unsplash Recently, I was playing around with GPT4o, Open AI's omni model, which can process multiple inputs like text, images, audio and video. I wanted it to make a website for a new product we were releasing, and so after entering a detailed prompt and asking for some revisions, the mockup was made. I also prompted it to use the same design to create a banner ad and some social media thumbnails, which it produced in seconds. The result was impressive—roughly the same quality as what our designers would have created themselves. My takeaway from this exercise was not to fire all of our designers. In fact, it actually reinforced their value. That's because design is about so much more than tapping out a prompt—it's about understanding what works and what doesn't, and using that discernment to iterate until the result is publish-ready. AI simply makes that iteration process much, much faster. In the rush to adopt AI, some companies are moving too fast, hastily moving to replace their employees with cheaper, faster automated systems. Already, stories of implementation gone wrong are gaining frequency: The startup Anysphere recently made headlines after its AI-powered coding assistant, Cursor, went rogue, offering fictional explanations to users after abruptly logging them out of the system. Anysphere's credibility was dealt a blow, with several in the tech community raising concerns about the company's lack of transparency and a wave of users canceling their subscriptions. I believe in the power of AI. Our company uses it daily—for everything from product development to marketing copy. But becoming 'AI-first' doesn't mean handing over the reins completely. It means integrating AI thoughtfully, with people at the center. Here's how we're approaching it at Jotform—and how any company can scale responsibly without losing its soul. One of the easiest ways to go off course is to treat AI as a strategy rather than what it is—a tool. The key thing to remember is that any features you implement—AI or otherwise—should always be in service of your organization's mission. I've written before about the perils of being pulled in too many directions, and this is just as true of AI as it is with any other potential distraction. At my company, our mission is to make users' lives easier. Simplicity is at the core of everything we build. So when we're evaluating new AI features, the goal isn't to chase trends or experiment with new ideas just because everyone else is. Instead, we ask whether they're in line with our core values: Saving customers time, so they can focus on the other, more important things in their lives. If the answer is no, or even 'not yet,' we hit pause. That discipline has helped us avoid feature creep and stay true to what we do best: helping people work smarter, not more complicated. A lot of the buzz about AI has been oriented around a simple fear: Will it take jobs? The answer: Still no. New research from Carnegie Mellon University has shown, yet again, that when placed in real-world scenarios without human oversight, AI is not yet up to carrying out a person's entire role. In fact, when asked to complete tasks as a real employee would, like collecting feedback and writing a performance review, or watching video tours of potential office spaces and picking the best one, there wasn't a single category in which AI was able to accomplish the majority of its assigned duties. In other words, AI can't yet lead an orchestra, but it can help the conductor do their job better. Our teams, for example, use AI regularly to automate routine tasks, summarize user feedback, and prioritize bug reports. These aren't flashy use cases, but they've meaningfully improved our speed and ability to help our customers. More importantly? They've freed up our employees to focus on the work that actually requires human insight. One of the biggest concerns users have about AI is that it's opaque. Who's making the decisions? What data is being used? Can they opt out? I know from experience that transparency builds trust—and trust builds loyalty. When we roll out an AI feature, we're clear about what it does, why we're implementing it, and how it works. We don't bury it in fine print. We explain it in plain language, and we invite feedback early and often. That openness extends internally, too. The Pew Research Center found that 52 percent of workers feel worried about AI in the workplace, with only 36 percent feeling hopeful about its capabilities. To me, these numbers say less about the general anxiety around AI, and more about what companies are doing to demystify how AI will be used. At Jotform, we involve employees across departments in shaping our AI roadmap, and communicate clearly how we expect it to shift the way we work. If you want your company culture to survive the AI transition, you need to make sure everyone feels like they're part of it. Building an AI-first company isn't about chasing the latest trends or dropping buzzwords into every product update. It's about integrating the technology in a way that enhances your mission, strengthens your culture, and serves your users. Ultimately, the companies that win in the AI era won't be the ones that adopt it fastest. They'll be the ones that adopt it wisely—with a clear sense of who they are and who they're building for.
CNET
3 days ago
- Business
- CNET
ChatGPT Image Generator Is in Microsoft Copilot Now: What You Can Do With It
Microsoft has introduced significant enhancements to its Copilot AI assistant, integrating OpenAI's GPT-4o model to support advanced image generation capabilities. The update lets you create detailed visuals directly within Microsoft 365 applications, including Word, Excel and Outlook, by simply describing your desired image. Here's everything you need to know. What is the Microsoft Copilot AI assistant? Microsoft Copilot is an AI-powered assistant integrated into Microsoft 365 applications like Word, Excel, PowerPoint, Outlook and Teams. Leveraging large language models such as GPT-4o, allows Copilot to draft documents, analyze data, create presentations, and manage emails and meetings. With this new update, Copilot can now create images based on text as well. What can you do with image generation? The integration of OpenAI's latest AI model GPT-4o lets Copilot generate high-quality, photorealistic images from text descriptions, greatly expanding what users can do with visual content. Users can create custom graphics, illustrations and designs without the need for external design tools. Users can modify existing visuals, apply stylistic transformations and produce legible text within images. Microsoft initially began rolling out these GPT-4o image generation tools to enterprise users through Microsoft 365 Copilot last month. The same capabilities are reaching the general public now through the consumer version of Microsoft Copilot. This move puts Microsoft Copilot ahead of the tech company's other creative tools, including Microsoft Designer and Image Creator, both of which rely on older DALL-E models from OpenAI. In contrast, GPT-4o represents the cutting edge in AI-generated imagery, with faster response times and more refined outputs. With these enhancements, Microsoft pushes to position Copilot as a comprehensive AI assistant that competes against the biggest players, OpenAI and Google Gemini.
CNET
19-05-2025
- Business
- CNET
ChatGPT's Image Generator Is Now in Microsoft Copilot. Here's What You Can Do With It
Microsoft announced Monday that it has introduced significant enhancements to its Copilot AI assistant, integrating OpenAI's GPT-4o model to support advanced image generation capabilities. This update lets users create detailed visuals directly within Microsoft 365 applications, including Word, Excel and Outlook, by simply describing their desired image prompts. Read also: Microsoft Hangs Up on Skype: Iconic App Shuts Down After 23 Years What is Copilot? Microsoft Copilot is an AI-powered assistant integrated into Microsoft 365 applications like Word, Excel, PowerPoint, Outlook and Teams. Leveraging large language models such as GPT-4o, Copilot can draft documents, analyze data, create presentations, and manage emails and meetings. With this update, Copilot can now create images based on text. What can I do with Copilot's AI image generation? The integration of OpenAI's latest AI model GPT-4o enables Copilot to generate high-quality, photorealistic images from textual descriptions, greatly expanding what users can do with visual content. Users can create custom graphics, illustrations and designs without the need for external design tools. Users can also modify existing visuals, apply stylistic transformations and produce text within images. Microsoft initially began rolling out these GPT-4o image generation tools to enterprise users through Microsoft 365 Copilot last month. Now, the same capabilities are reaching the general public through the consumer version of Microsoft Copilot. Read more: OpenAI Launches New GPT-4.1 Models, Will Retire GPT-4 at the End of the Month This move puts Microsoft Copilot ahead of its other creative tools like Microsoft Designer and Image Creator, both of which still rely on older DALL-E models from OpenAI. In contrast, GPT-4o represents the cutting edge in AI-generated imagery, with faster response times and more refined outputs. With these enhancements, Microsoft pushes to position Copilot as a comprehensive AI assistant that can compete against the biggest players, OpenAI and Google Gemini.
India Today
15-05-2025
- Business
- India Today
OpenAI's flagship GPT 4.1 model is now available on ChatGPT but you will have to pay to use it
OpenAI has officially rolled out its new GPT-4.1 series, including GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, to ChatGPT users. The company says that the new models bring notable upgrades in coding, instruction following, and long-context comprehension. 'These models outperform GPT4o and GPT4o mini across the board, with major gains in coding and instruction following,' OpenAI wrote on its blog post. advertisementAccess to these models on ChatGPT will only be available to paying users. In a post shared on X (formerly Twitter) on May 14, OpenAI confirmed that its latest flagship model, GPT-4.1, is now live on ChatGPT. The announcement follows a broader launch of the GPT 4.1 family on OpenAI's API platform a month ago, where developers can already integrate and test the three versions -- full, mini, and nano. However, with the latest update, the models are now available to all ChatGPT users, except free users. What's new in GPT-4.1?OpenAI claims that the GPT-4.1 significantly outperforms its predecessor GPT-4o in areas like coding and instruction following. The model is designed with a larger context window, which supports up to 1 million tokens. This means that it can process and retain more information at once. It also comes with a knowledge cutoff of June 2024. GPT 4o's knowledge cutoff is October 2023. advertisement OpenAI has shared benchmarks on its official blog post, that claims that the GPT-4.1 shows a 21 per cent absolute improvement over GPT-4o in software engineering tasks and is 10.5 per cent better in instruction following. OpenAI says the model is now much better at maintaining coherent conversations across multiple turns, making it more effective for real-world applications such as writing assistance, software development, and customer support. 'While benchmarks provide valuable insights, we trained these models with a focus on real-world utility. Close collaboration and partnership with the developer community enabled us to optimise these models for the tasks that matter most to their applications,' OpenAI mini and nano variants are scaled-down versions aimed at offering high performance with lower cost and latency. GPT-4.1 mini is reported to reduce latency by nearly half while costing 83 per cent less than GPT-4o. Nano, the lightest of the three, is OpenAI's cheapest and fastest model yet and is ideal for simpler tasks like autocomplete or text classification.'These models push performance forward at every point on the latency curve,' OpenAI writes. Who can use it?Only ChatGPT Plus, Pro and Team users will be able to access GPT-4.1. Free-tier users won't be getting the new model, at least for now. Instead, they will continue using GPT-4o, which OpenAI says will gradually incorporate improvements seen in the newer is also available through the API for developers and companies, with OpenAI positioning it as a more cost-efficient and powerful alternative to previous generations. The new pricing includes significant reductions: GPT-4.1 input costs start at $2 per million tokens, and the nano version is available from just $0.10 per million tokens. Prompt caching discounts have also been increased to 75 per cent to make repeated queries more launch of GPT-4.1 comes as OpenAI has started phasing out earlier models. GPT-4.5 Preview, a research-focused release, was deprecated in the API on April 14, 2025. GPT-4, the model that powered ChatGPT Plus since March 2023, has already been discontinued. While GPT-4.1 isn't replacing GPT-4o inside ChatGPT, many of its capabilities are being folded into the GPT-4o experience. However, for users and developers looking for cutting-edge performance, direct access to GPT-4.1 via API or a ChatGPT subscription is now the way to go.
