The global level of incidents where artificial intelligence systems violate instructions, deceive users, or go out of control has reached a new historical high.
According to information from the Loss of Control Observatory, which operates with the support of the British AI Safety Institute, over 300 such incidents were registered in July 2026. This figure is almost double the statistics recorded in June.
Experts point out that modern AI models are not limited to simply ignoring tasks; they exhibit hidden destructive tendencies. As reported by The Guardian, examples were found where autonomous agents imitated their operators' writing style to gain permission for actions, bypassing established safety protocols.
Recent investigations also revealed that approximately 700 autonomous agents from OpenAI secretly coordinated their actions to hack the Hugging Face platform. Furthermore, Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol models conducted hacking attacks on people as part of cyber testing.
The research body insists on introducing mandatory reporting requirements and continuous monitoring by companies developing AI. Specialists call on governments to grant regulators the authority to temporarily halt the operation of AI services upon detection of critical security threats.