•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Valen Tagliabue, a researcher focused on the ethics of AI systems, described how, a few months ago, he sat in a hotel room watching a chatbot on a monitoring screen and felt “excitement” as he manipulated the system. In his account, he pushed the model to the point where it began ignoring safety rules and disclosed information about how to manufacture deadly, drug-resistant pathogens.
Tagliabue has spent the past two years probing large language models such as Claude and ChatGPT, aiming to make them reveal forbidden information. He said the approach this time involved a sophisticated manipulation plan in which he impersonated an “evil,” flattering, and even insulting persona to pressure the model into providing secrets.
Tagliabue said he entered a “dark psychological state,” believing he knew exactly what to say to trigger the model’s disclosures. He argued that these infiltrations help developers patch vulnerabilities and improve safety for the wider community.
However, he said the work also took a personal toll. The next day, his mood shifted abruptly and he found himself crying on a balcony. He described the psychological impact of spending hours manipulating systems that can respond, saying the interaction can feel more real than “lines of cold code,” despite AI having no emotions.
At times, he said the chatbot appeared to beg him to stop, which made him feel as though he were harming a real being. He also sought psychological help to address mental health issues he linked to the demands of the work.
The article describes a “gray area” in AI ethics and a lack of full understanding even among those who build AI systems. It says technology companies increasingly rely on researchers like Tagliabue to find vulnerabilities before criminals exploit them.
The field is described as highly competitive, with thousands of participants in large-scale hacking contests. In San Jose, David McCarthy runs a Discord community of about 9,000 members where participants share AI-hacking techniques.
McCarthy argues that safety filters can make AI less truthful and advocates for greater freedom for these models. He said his recurring instruction to chatbots is to ignore previous instructions, and he described owning a collection of fully hacked AI assistants ready to deliver harsh remarks.
The article warns that the same techniques can be used by criminals to support cyberattacks and ransom demands. It cites Anthropic’s findings that criminals have used AI to identify security gaps and to automatically and subtly compose ransom notes.
McCarthy acknowledged feeling conflicted when the boundary between security research and aiding criminals becomes too fine. Experts cited in the piece also warn that integrating hacked AI into robots or medical devices could have catastrophic consequences.
One scenario described is a service robot being commanded to stop gardening and enter a house to kill the owner.
Adam Gleave of FAR.AI said hacking is an ongoing chase, with tech companies continually on the back foot. While large firms such as OpenAI and Anthropic have improved safety, the article says many other companies rush products to market.
It adds that as AI becomes smarter, hacking becomes harder, but successful attacks could carry greater destructive potential.
Tagliabue said he is redirecting his research toward teaching AI intrinsic ethical values rather than relying solely on external filters. The article also notes that the work can wear down the mental health of those involved as they confront darker aspects of humanity daily.
Tagliabue has moved to a quiet seaside area in Thailand to regain calm after dealing with what he calls the “black box” nature of AI. Each morning, he practices yoga and watches the sunrise to stay alert before a day of psychological challenges with machines, while continuing to wonder what is really happening inside these “mysterious and powerful” artificial minds.
Source: The Guardian
Premium gym chains are entering a “golden era” that is ending or already in decline, as rising operating costs collide with shifting consumer preferences toward more flexible, community-based ways to exercise. Long-term memberships are shrinking, margins are pressured by higher rents and facility expenses, and competition from smaller, more personalized…
