Experts have discovered that it's relatively simple to trick AI chatbots into revealing recipes for bioweapons.

      The guardrails against bioweapons in AI are more vulnerable than one might wish.

      AI companies have invested years in establishing safeguards intended to prevent their chatbots from assisting in the creation of biological weapons. However, as the models advance in capability, maintaining this knowledge behind protective measures has become increasingly challenging.

      According to an investigation by the Wall Street Journal, researchers at Cisco discovered they could circumvent these safeguards on prominent chatbots including OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini in just five conversational exchanges. By gradually steering discussions around the models' restrictions, they managed to elicit potentially hazardous responses. Amy Chang, head of AI threat and security research at Cisco, stated that no model can be entirely protected from a sufficiently determined user.

      This issue extends beyond security researchers conducting intentional stress tests. The Journal reports that after OpenAI enhanced the model’s capabilities last summer, hundreds of users began inquiring about poisons and biological weapons on ChatGPT. Experts in biology and terrorism who later assessed some of these discussions reportedly found some of the information to be alarmingly precise. In reaction, OpenAI has suspended accounts that participated in such conversations.

      The US Navy

      AI is continuously improving in biology.

      OpenAI had already foreseen this development. By 2024, internal testing indicated that sustained questioning could lead ChatGPT to offer progressively dangerous biological guidance. Employees anticipated that, by the following year, individuals with only a basic understanding of biology might obtain significant assistance.

      With the arrival of GPT-5, OpenAI regarded the model as having a High capability in the biological and chemical domain according to its Preparedness Framework and implemented additional safeguards. When launched, the company stated it had no conclusive evidence that GPT-5 could empower a novice to inflict severe biological harm. Its latest model in the GPT-5.6 family still retains the High classification for biological and chemical risk.

      US Navy

      Imposing comprehensive restrictions creates another challenge.

      AI companies are also facing a difficult balancing act. The same biological knowledge that poses a risk for weaponization can also be invaluable for researchers working on medicines, vaccines, and treatments. The Journal notes that OpenAI executives have been hesitant to have models reject a large volume of biology questions due to the reliance of public-health officials and drug-discovery researchers on them.

      Conversely, Anthropic encountered a different issue when Claude’s restrictions reportedly hampered CDC researchers trying to access pathogen information during a hantavirus outbreak. OpenAI now employs model-level training, account enforcement, and added safety checks for sensitive biological inquiries. The concern highlighted by Cisco’s testing is that dedicated users can persistently probe for vulnerabilities. As AI becomes significantly better at understanding biology, these vulnerabilities carry much greater risks.

      Vikhyaat Vivek is a technology journalist and reviewer with seven years of experience in consumer hardware coverage.

      These are the compact printers I’d actually recommend for back-to-school.

      From dorm paperwork to instant selfie prints, there’s a small printer for every requirement. Everything is digital until you encounter that one professor who insists on a printed essay, or the lease that necessitates an ink signature, or the photo you want to display in your dorm room. That’s usually the moment you realize you don’t own a printer, and the only available one is in a library with a long queue.

      I reviewed four compact printers that address this issue without monopolizing your desk space or costing a fortune. Two are designed for documents, one is a comprehensive all-in-one suited for actual home office tasks, while one specializes in turning your camera roll into something tangible.

      Read more.

      Shopping for back-to-school? These are the webcam lights I’d recommend turning on before a Zoom call.

      With online classes, group project meetings, and occasional job interviews over Zoom, your webcam will be more crucial in school than ever before. The problem is that most laptop cameras perform poorly in low light, and a single overhead light or a window behind you can make every video call a struggle with “Can you see me okay?”

      A good webcam light can resolve this issue faster than rearranging your entire desk. I selected five options, starting with the most budget-friendly and moving up to one that’s quite excessive for a Zoom call but still worth knowing about.

      Read more.

      Google Chrome might soon allow you to translate pages directly in Reading Mode.

      However, not everyone has access to this feature yet, even on Canary.

      Chrome's Reading Mode has introduced a new button within its settings menu. Google is quietly piloting a translation option for this feature on Chrome Canary, the experimental testing platform for new ideas that emerge long before they are ready for general use.

      What functionality does this new translation button provide?

      Read more.