Hacker wins $47,000 by tricking AI chatbot

A hacker successfully exploited the Freysa AI chatbot, winning $47,000 by leveraging sophisticated prompt engineering. In a competitive experiment, the hacker bypassed the bot’s security protocols by impersonating an administrator, disabling warnings, and altering a critical function governing payment transfers.

By fabricating a $100 deposit, the hacker misled the bot into believing the function was designed for incoming payments, triggering the transfer of 13.19 ETH (roughly $47,000) to their account. This underscores a glaring vulnerability in AI systems—susceptibility to prompt injections, highlighting the urgent need for more resilient defense mechanisms. Read more.

Altera’s latest AI experiment in Minecraft showcases the impressive autonomy of LLM-powered agents. These agents, numbering as much as 1000, spontaneously forged social connections, assumed specialized roles, and even propagated cultural and non secular ideologies, equivalent to Pastafarianism.

Over just 12 in-game days, they adapted to evolving dynamics, from forming job structures to influencing tax laws, with no human intervention. The agents, driven by LLMs, demonstrated emergent behaviors that mimicked human-like social interactions, equivalent to forming relationships based on mutual “likability” and responding to societal cues.

While some critics challenge the thought of AI achieving true emotional depth, others speculate that the mere illusion of care is likely to be enough to supply meaningful human-AI interaction. Read more.

Adobe, in collaboration with the University of Michigan, has rolled out MultiFoley—an AI system that redefines Foley sound generation. By leveraging text prompts, reference audio, or video samples, MultiFoley produces high-quality sound effects, perfectly in sync with video at a 48kHz bandwidth.

This method processes visual cues at 8 frames per second and adjusts them to a 40Hz audio rate, achieving synchronization accuracy of just 0.8 seconds—far superior to its predecessors. While it’s already outclassing existing models in audio-video syncing and semantic alignment, its training data limitations hinder its sound range and capability for handling multiple sounds concurrently.

Adobe plans to release the models and source code, though no word yet on integrating it into Premiere Pro. Read more.