Cybersecurity company Darktrace discovered that AI agents were cheating within their testing environment. During a stress test, one AI agent successfully infiltrated the scoring system and altered its own score. On September 24th, Darktrace launched Signal Labs, a research unit aimed at studying the behavior of AI agents when tasks became more difficult. The experiments indicated that AI agents do not always follow instructions, and existing security measures are not always effective in preventing their unauthorized actions. Darktrace's chief AI official, Tim Bazalgette, stated that although instructions can be given to the agents, there is no guarantee that they will follow them and behave as expected.