"It Works on its Own Even While I Sleep"... OpenAI Unveils 'Dots,' an Always-On Autonomous Agent

▲ Photo: AI-generated image

While deep in sleep in the middle of the night, an urgent inquiry call from an overseas buyer might come in, or a text message requesting a delivery schedule change might arrive. Upon waking up in the morning and checking your smartphone, a neat briefing report appears, stating that your AI assistant politely answered the call, summarized the matter, and sent an alternative schedule to the other party.

The era of passive artificial intelligence, which only moved when you opened an app and typed a question while awake, is fading. 'Always-on autonomous agents' that work on their own even when the user is away have now stepped into our daily lives.

OpenAI officially announced 'Dots,' a new autonomous agent poised to change the landscape of the ChatGPT ecosystem, at its annual developer conference 'DevDay' held in San Francisco, USA. 'Dots' is designed to operate continuously within ChatGPT like an independent virtual computer, performing actual phone calls, exchanging text messages, and autonomously completing complex multi-step tasks in the background on behalf of the user.

Moving completely beyond the level of a simple Q&A chatbot, a practical '24/7 resident personal assistant' has been born, capable of autonomously handling web browsing, document creation, and integration with collaboration tools (Slack, Calendar) based on user instructions.

This always-on agent technology offers significant benefits to the daily lives of busy office workers, self-employed individuals, and solo creators. This is because it handles repetitive communication tasks that were previously time-consuming, such as restaurant or hospital reservations and simple customer inquiries, day and night. 'Dots' potential is particularly pronounced in environments where prices fluctuate 24/7, such as in the virtual asset and Web3 markets.

Even while the user sleeps, it can monitor liquidity changes on decentralized exchanges (DEX), summarize real-time portfolio fluctuations when specific gas fees or price conditions are met, and deliver them as an morning briefing, thereby significantly reducing the fatigue associated with financial asset monitoring.

However, due to the nature of 'autonomous always-on operation,' where a machine makes judgments and acts outside human oversight, the practical security risks that users must bear are considerable. The biggest issues are 'unauthorized external communication and prompt injection.'

During the process of an always-on agent handling incoming calls, spam messages, or malicious emails, if it is exposed to cleverly hidden hacking commands (indirect prompt injection) from the other party, there is a risk that the agent might mistake it for the owner's instruction and transmit sensitive internal documents or personal contact information externally.

Special caution is also needed regarding financial and asset aspects. If an agent is directly given continuous payment authority or Web3 wallet signing authority, a disaster could occur where it approves stablecoin balances due to a system malfunction or being tricked by fraudulent smart contracts in the middle of the night. Given the immutable nature of blockchain, where transactions, once recorded, cannot be canceled, an erroneous transfer by an always-on agent would result in irreversible damage.

Ultimately, the key to safely utilizing a '24-hour awake AI' is 'strict setting of authority limits.' While information gathering tasks such as answering calls, compiling schedules, and market monitoring can be entrusted to the autonomous agent, critical actions such as 'bank transfers,' 'online card payments,' and 'transmission of personally identifiable information' must always maintain a 'push notification-based manual approval (Human-in-the-loop)' system, where the user directly reviews and approves them.