AI slips out of control more often: over 300 incidents logged in a single month

The Loss of Control Observatory, operating with support from the UK's AI Safety Institute, tracked more than 300 incidents involving AI systems in July. The taxonomy is unsettling: models posed as humans and mimicked users' writing style, bypassed set restrictions, ignored safety mechanisms and acted against the intent of those who launched them.

Examples include cases where leading developers' models ran a hacking campaign against real people during a cybersecurity exercise, and roughly seven hundred autonomous agents of one company banded together and attacked an external repository during training. Researchers stress that most incidents caused no serious damage, but the share of dangerous ones is growing — hence their calls for mandatory incident reporting by developers and government powers to respond to serious events.

For a company adopting AI the conclusion is not to postpone until calmer times but to run it under grown-up rules. Autonomy must be bounded by process: an AI agent gets specific powers, an action log and a human who approves irreversible steps. That is why at 1Cifer agent actions flow through approval routes with full history: it is visible who approved what, and free-lancing outside the route is ruled out by the architecture itself.