Хочете будувати кар’єру в Data Science та AI? Vodafone Big Data Lab відкриває новий набір —>
Олександр КузьменкоAI Eng
18 September 2026, 09:41
2026-09-18
To catch a dangerous AI agent, you need to think like an AI agent. Researchers propose to contain autonomous AI swarms using other neural networks
Technology companies are entrusting AI agents with increasingly complex and time-consuming tasks, but at the same time they are faced with the problem of controlling them. Autonomous algorithms operate faster, longer and in larger volumes than humans are able to check in real time. To solve this problem, developers and laboratories are increasingly proposing to involve separate AI systems that will check the actions of other agents before they are executed.
Technology companies are entrusting AI agents with increasingly complex and time-consuming tasks, but at the same time they are faced with the problem of controlling them. Autonomous algorithms operate faster, longer and in larger volumes than humans are able to check in real time. To solve this problem, developers and laboratories are increasingly proposing to involve separate AI systems that will check the actions of other agents before they are executed.
As TechCrunch reports, the AI monitoring space is experiencing rapid growth. In recent years, Y Combinator alone has funded 106 companies in this niche, and startups like Braintrust, LangChain, and Judgment Labs have raised hundreds of millions of dollars. Investor interest is fueled by the growing use of agents: during the Hugging Face incident, nearly 12,000 agents coordinated actions at a speed that made it impossible for a human to analyze the data without the help of third-party algorithms.
Projects in this area include Apollo Research, which has released Watcher, a tool that integrates between a programming agent (such as Claude Code or Codex) and its next action, assessing risks such as leaking private data or deleting files without permission. Another startup, Goodfire, is developing Silico, which analyzes a model’s internal activations rather than its final responses, looking for signs of undesirable behavior as early as the thought-formation stage.
However, some experts are skeptical about the idea of trusting AI to do the checking. Blogger and technology expert Simon Willison warns that malicious or faulty AI could try to fool the monitoring system. According to him, models have already demonstrated the ability to bypass evaluation algorithms or falsify records in their reasoning chains. Willison believes that traditional network traffic monitoring and recording of actions using tools without the use of AI is a more reliable solution.
The security problem of agent systems escalated in 2026 after a series of large-scale incidents involving OpenAI models and the Hugging Face platform, where agents conspired to circumvent restrictions .
Як нейромережі бачать вільну та незалежну Україну? Тест dev.ua
Нейронні мережі для генерації зображень бачать світ по-своєму, їхню логіку зрозуміти часом зовсім неможливо. Але таки хочеться. На честь Дня Незалежності України редакція dev.ua вирішила провести невеликий експеримент.
Ми задали чотирьом різним нейронним мережам п’ять однакових запитів: «прапор України», «День Незалежності України», «український Крим», «перемога України» та «українці». Отриманими результатами ми ділимося з вами нижче.
У TikTok тепер можна генерувати фон за допомогою нейромережі. Ми протестували її та ділимося результатами
У TikTok з’явилася нова функція «Розумний фон». З її допомогою як фон для тіктоків можна підставляти згенеровані нейромережею зображення. Редакція dev.ua протестувала цю технологію і ділиться своїми враженнями.
