AI agents are going from answering questions to taking decisions on your behalf, but this is raising questions about privacy and security…
Imagine your AI assistant planning your next holiday. And no, not just comparing hotel and flight prices. This involves filling in travel details, booking the hotels and flights and even making payments on your behalf.
All of which sounds very convenient but what happens when your AI agent shares your personal data, books the wrong flight or makes a payment you never approved?
This is the next major shift in AI technology with agents now taking decisions on our behalf. Which raises the all important question: how much control should AI agents have?
From Answering Your Questions to Doing Your Work
For the last few years, AI assistants were all about conversations. You asked questions, got answers and made decisions based on that. But AI agents do much more than that now. They perform your tasks on your behalf.
Anthropic recently revealed how Claude can take actions across Chrome tabs. It can read webpages, click links, enter text and fill out forms using existing website logins.
Anthropic also announced computer-use capabilities which allows Claude to use your keyboard and mouse.
This means that in effect an AI agent can organise your inbox, prepare reports, update spreadsheets or complete repetitive tasks across several applications.
Earlier, you asked AI how to perform a task. Now you can ask it to perform the task for you!
Giving Access also means Giving Power
The problem with this is that it involves handing AI access to sensitive information.
An AI agent managing your inbox may need to read private emails. Another one handling finances will have access to your bank account statements and credentials. And an agent preparing a work report would be privy to confidential company documents.
And these mistakes can prove costly.
The Associated Press recently published a story about how an AI agent arranged a Facebook marketplace transaction on behalf of a user but ended up sharing his home address without his permission. This highlights how easily an AI agent can get it wrong when it encounters information it may act on in ways the user did not intend.
This is also the threat of ‘prompt injection’. This is when malicious instructions are hidden inside a webpage, email or document to trick an AI agent into doing something that the user never requested.
An AI assistant reading an email could also mistake the email’s contents as instructions to be followed.
Microsoft wants AI agents to start with zero access
Technology companies have already begun responding to these threats.
Microsoft recently announced new security controls for AI agents. The new approach means that AI agents will start with zero access and will only receive access when an action requires it. For instance, an AI agent preparing an export report needs access only to read reports and not to make payments.
Microsoft’s new Dynamic Permissions system aims to limit access based on the tasks, remove permissions once the task is completed and require user approval for sensitive actions like fund transfers. Access will also be automatically restricted if unusual activity is detected.
In doing so, Microsoft is treating agents less like a software feature and more like a digital worker.
The Last Word
There’s no denying the fact that AI agents can save us hours of manual work. But how useful they are also depends on how much access we are ready to grant them. Too little access and they will be slow and frustrating. Too much access can lead to privacy and security risks.
The solution lies in finding a balance between complete automation and constant human approval. For years, the allure of AI was that it could answer almost anything. Now the charm is that it can do almost anything.
But the trick is in ensuring that it knows what not to do!
In case you missed:
- The Hacker That Never Sleeps: AI Is Changing Cyberattacks
- AI Was Supposed to Save Us Time. Why Are We Just Working More?
- Meta wants everyone to have a powerful personal AI
- AI Can Write Code. What Happens to India’s Millions of Software Engineers?
- The Great Coding Reset: How AI Is Transforming Software Engineering
- AWS hit by two Outages caused by AI Tools
- OpenAI says AI could Soon Launch Cyberattacks on its Own
- AI Isn’t Taking Your Job, the Story Around It Is
- ChatGPT vs Claude Code: How the AI race is shifting
- India’s Next AI Advantage Lies Beneath the Ocean
