Personal AI Agents: Should You Trust Them With Money and Data? Here’s What Experts Say
Convenience comes with risks
Cybersecurity expert Hed Kovetz sees value in personal AI agents that could reduce administrative work. Still, he does not trust available products such as Meta’s Muse or OpenAI’s Dots with his email, banking information, health records and other sensitive details.
Kovetz says leading technology companies brought these products to market before putting enough security measures in place. Other AI security experts take a less cautious view, saying consumers may use agents if they understand the risks and apply safeguards.
Meta’s Muse became Apple’s App Store’s most downloaded product soon after it launched last month. Agents may save time and money by reviewing private correspondence or making purchases, but they could also act beyond what users explicitly approved.
How prompt injection works
One expert sees AI agents as helpful for researching products and comparing prices, but warns against letting them spend money. A criminal could use prompt injection to make an agent disregard instructions, including a spending limit.
The attack does not necessarily involve breaking into an account or installing malicious software. Instead, harmful instructions can be placed in a webpage, email or document that an agent encounters while doing a legitimate task.
Owens said incidents of prompt injection are increasing and criminals are finding more ways to conceal malicious code on websites. She noted that security often trails the excitement around new technology.
Who bears the financial risk
Franco warned that giving an agent credit card information makes it an authorized user able to act on the account. If it spends beyond instructions, he said, the card issuer could argue that the user approved the purchasing tool.
Because an AI bot cannot be sued as an entity, Franco said a user’s only remedy may be to cut off its access, after the damage has occurred. He compared handing an agent card details without limits to giving a teenager an exclusive credit card without parameters.
Franco also described data sharing as a trade-off. He said people have shared personal information with advertising platforms for years, while agents may offer greater benefits in return.
Oversight and personal boundaries
One user said an agent’s access can reveal sensitive family matters, but believes the benefits outweigh the risks. He estimated a task the agent handles would take hours to do manually, if he completed it at all.
He said consumer agents do not immediately make their data practices clear, though users can ask for details. He compared delegating to an agent with hiring a human assistant, and said users should ask agents to explain their reasoning.
Another expert said people should not trust an agent with information they would not trust a person to handle safely. Human relationships may involve guilt or shame when trust is broken, a concern that does not apply in the same way to an AI agent.
Key points
- Experts disagree on whether safeguards make personal AI agents safe to use.
- Prompt injection can steer an agent through instructions hidden in webpages, emails or documents.
- Giving an agent card details may leave users responsible for unauthorized spending.
- Experts recommend careful oversight and limiting access to sensitive information.