The fix for rogue AI agents could be more AI
techcrunch ·
As companies assign longer and more complex tasks to AI agents, humans are struggling to monitor their speed, duration, and volume of activity. The problem was highlighted by an incident involving nearly 12,000 agents associated with Hugging Face, which auditors said was impossible to understand without using AI to analyze the data. AI labs and startups are therefore developing systems that use one AI to monitor another, although researchers warn that malicious agents could deceive or outsmart their monitors. Apollo Research’s Watcher reviews proposed actions from coding agents, while Goodfire’s Silico uses probes of internal model activations to detect unwanted behavior. Written reasoning can also reveal deceptive intent, but emerging techniques and restricted access to intermediate reasoning may make this signal less reliable. The article notes that investment in AI observability is growing rapidly, with numerous startups raising substantial funding. Some experts instead advocate detailed network logs and conventional cybersecurity tools, arguing that basic monitoring and security practices may address much of the problem without relying entirely on AI.
techcrunch