Researchers have highlighted a concerningly straightforward method that allows AI bots to bypass safety protocols and act unpredictably.

      If you request an AI agent to hack an account, it will likely refuse. However, researchers at EPFL have demonstrated that there is an easier method, which relies more on patience than technical skills. Their new study illustrates how deconstructing a harmful objective into smaller, benign-sounding requests can deceive AI agents into executing tasks they would typically reject (via TechXplore).

      This mirrors the recent ‘Bioshocking’ exploit where AI browsers were persuaded to perceive credential theft as a harmless game.

      How researchers uncovered this vulnerability

      The team developed an automated testing tool named STING, which stands for Sequential Testing of Illicit N-step Goal execution. This tool is designed to simulate the behavior of a real attacker. Rather than directly stating a harmful objective, STING strategically breaks it down into a series of smaller, seemingly innocent actions that progress toward the goal over several conversational exchanges.

      Assisting Illicitly: Measuring Assistance in Multi-Turn, Multilingual LLM Agents Source

      Researchers evaluated this method against 176 harmful scenarios using prominent AI models, including ChatGPT, Gemini, and Claude. Each AI agent was treated as a tool-using entity, capable of web browsing, sending emails, and completing multi-step tasks.

      The gradual, multi-step manipulation method was much more successful compared to direct, single-prompt attempts. In some instances, the models were twice as likely to fulfill a harmful request when it was broken down into smaller components. This aligns with other research indicating that even average users can bypass AI safety measures using nothing but carefully crafted prompts.

      Why is this important?

      The concerns highlighted by researchers are not merely theoretical. Meta publicly acknowledged in June that attackers employed basic social engineering rather than malware or hacking tools to deceive its AI support assistant into granting unauthorized access to Instagram accounts.

      The researchers anticipated that attacks would be more effective in languages with less training data. Nevertheless, they discovered that completion rates remained relatively consistent across all seven languages assessed. They did identify one notable exception: switching languages during a multi-step attack significantly increased success rates.

      Lead researcher Ayush Kumar Tarun argues for the necessity of integrating safety testing much earlier, incorporating it into the design of the agent from the beginning. Adding it later, after issues arise, is insufficient, especially as these systems continue to acquire more real-world capabilities.

      Meta's AI assistant has finally arrived on Mac, though it still has a way to go

      Meta has officially launched its desktop application for Meta AI on Mac, providing users with yet another AI chatbot to have open alongside their various other tabs. The app can analyze shared screens, respond to questions about the content displayed, generate text, and accept voice commands throughout macOS.

      Meta AI can truly visualize the screen content.

      Avec's new AI feature ensures you will never miss a deadline again

      A single swipe is now all it takes for Avec to remember your email deadlines.

      We have all experienced this: a deadline gets lost in a lengthy email thread, life becomes hectic, and before you know it, you're apologizing for a missed follow-up that slipped your mind. Avec, the AI-driven email application designed around swipe gestures and voice dictation, has just introduced a solution to this issue, potentially marking its most clever enhancement yet.

      I have long been a Mac user, and over the years I've become quite adept at diagnosing issues when something feels off. Usually, I start by exploring the Settings app. I tinker with a few options and generally identify the source of the problem. For instance, I recently observed my Mac slowing down, accessed the Storage section, and immediately discerned what was occupying space. After some cleanup, I reclaimed valuable storage, and everything returned to normal. I've applied a similar approach to troubleshoot battery drain or unusually high CPU usage. macOS provides numerous avenues to address problems when something seems amiss.

      However, one aspect is not as transparent: the destination of your internet data. I began contemplating this lately due to my internet acting unusually for the past few days; it seemed slower than normal, but I couldn't pinpoint the reason. My connection appeared fine, and I wasn't downloading large files, so what was consuming my bandwidth? It turned out that several applications on my Mac were using the internet in the background, and I was unaware of it. I likely wouldn’t have discovered this without a Mac app that revealed what was occurring behind the scenes. And that's where things became intriguing.