Here's what you need to know
Cybersecurity experts caution against giving agents access to your personal information
Imagine a chatbot that can not only recommend a great restaurant in your area, but can also book a reservation, put it in your calendar and email an invitation to your friends.
asserted
that → need → friends
In recent months, that scenario has become reality, with some of the world's biggest AI developers launching digital personal assistants they say will make users' lives easier.
asserted
lives → become → assistants
The companies behind these new mass market agents say they are remarkably capable, safe and easy to use.
asserted
they → say → agents
But cybersecurity experts say the more advanced tasks come with more risk that something could go seriously wrong.
uncertain
something → say → risk
What are AI assistants, and how do they work?
asserted
they → work → ?
Three of most well-known ones on the market — Meta's Muse, OpenAI's dots and Anthropic's Claude Duet — all debuted within weeks of each other and all have similar capabilities.
asserted
all → know → capabilities
Muse can be spoken to as if it were a chatbot and instructed to send emails, book travel, make purchases and more on your behalf when connected to third-party apps like Shopify and Ticketmaster.
asserted
it → speak → Shopify
It can handle multiple tasks at once and continue working through them even while the app is closed.
asserted
app → handle → them
OpenAI says users can instruct dots to do "nearly anything" by messaging or calling them through ChatGPT.
asserted
users → say → ChatGPT
They can be connected to other devices and over 4,000 apps, and even have the ability to use your laptop.
asserted
They → connect → laptop
Anthropic's Claude Duet has been pitched as more of a proactive agent for work than as an assistant to take on personal errands.
asserted
Duet → pitch → errands
The agent can take on simple administrative tasks, like scheduling meetings, and more complex projects, like creating financial reports or building web apps, according to the company's product announcement.
uncertain
agent → take → announcement
While Muse is free, dots and Claude Duet each require a subscription.
asserted
dots → require → subscription
How do AI personal assistants keep your personal info safe?
asserted
assistants → keep → info
CBC News reached out to Meta, OpenAI and Anthropic to ask how they protect users' privacy and prevent their personal assistant agents from going beyond the scope of the tasks they're assigned, along with what advice they'd give users to prevent this from happening.
asserted
they → reach → this
A spokesperson for Meta said Muse runs on its own dedicated computer in the cloud that is designed so users' data is securely stored, referring CBC to a blog post by software engineer Tarek Sheasha outlining the safeguards built into Muse.
They also referred CBC to another blog post that says Muse checks with users before taking on sensitive actions, like sending an email or making a purchase, and shows a complete audit trail of everything the agent has done.
asserted
agent → say → everything
People are using AI to 'dumb down' their smart TVs.
asserted
People → use → TVs
Google launches AI-based shopping in Canada in race to dominate e-commerce
asserted
Google → launch → commerce
A spokesperson for OpenAI referred CBC to a blog post outlining its safeguards, which includes encryption for users' content and account security systems that monitor for suspicious use.
asserted
that → refer → use
OpenAI also monitors dots as they plan and carry out actions "looking for potentially harmful behavior, such as acting outside your instructions or trying to bypass safeguards," the blog post says.
asserted
post → monitor → safeguards
OpenAI also notes that users can give dots custom rules, such as to never send emails, and says certain actions require a user's confirmation each time, such as permanently deleting data and installing or running software from an unrecognized source.
asserted
actions → note → source
A spokesperson for Anthropic said Claude Duet was built to enable people to stay in control of what it does, and so their information is handled with discretion.
asserted
information → say → discretion
The model never sees users' passwords, browser cookies or sign-in tokens, and Claude Duet is instructed never to share sign-in credentials, credit card and bank account numbers and government ID numbers — even if you ask, the spokesperson said.
asserted
spokesperson → see → credentials
What can go wrong?
asserted
What → go → ?
Some users have reported getting more than they bargained for when putting these high-powered assistants to work.
asserted
they → report → work
A Montreal man asked Muse to handle his Facebook Marketplace transactions.
asserted
man → ask → transactions
But after listing a keyboard for sale, Muse gave out his home address to a stranger without his knowledge.
asserted
Muse → list → knowledge
When the buyer showed up, Muse even told him the seller was there even though he wasn't home.
asserted
he → show → him
"Muse doesn't check with me, it doesn't tell me this guy has messaged and then it gives the guy my address, agrees to his price and literally tells him to come right now, that's the crazy part," Matt Robb told CBC News last month.
asserted
Robb → check → News
Asked how something like this could happen, a spokesperson for Meta pointed CBC to a social media post made by Robb about his ordeal, where he explained he had clicked "allow always" instead of "allow one time" when asking Muse to handle his Marketplace transactions on his behalf going forward.
uncertain
he → ask → behalf
While these agents can be innovative and exciting tools, they should be used in a limited way and given very specific instructions in order to avoid these kinds of mishaps, said Claudiu Popa, a Toronto-based certified cybersecurity professional with privacy consulting firm Datarisk Canada.
asserted
Popa → use → Canada
"Unless you tell them very specifically what they can and cannot do you're going to have unpredictable results," he said.
asserted
he → tell → results
"That's the issue with all AI agents, and this issue is going to continue for the foreseeable future."
asserted
issue → go → future
David Shipley, CEO and co-founder of Beauceron Security in Fredericton, said he hasn't let these agents get near his computer at all.
asserted
agents → say → computer
"I absolutely do not trust this technology," he said.
asserted
he → trust → technology
Shipley compared trusting them with your personal errands to a monkey's paw wish, referring to the classic horror story that serves as a cautionary tale to be careful what you wish for.
"
asserted
you → compare → what
You might wish for it to do a certain set of actions and it's going to go and do those actions, but perhaps not in the way that you intended it to do that."
uncertain
it → wish → that
Why have Canada's biggest banks blocked AI assistants?
asserted
banks → block → assistants
With a client's authorization, AI agents can also connect to bank accounts and credit cards to monitor balances and transaction histories and give tips on budgeting.
asserted
agents → connect → budgeting
But concerns over AI agents accessing sensitive information have also prompted Royal Bank, TD Bank, CIBC and Bank of Montreal to block them from directly accessing their clients' online bank accounts and other information.
asserted
agents → access → accounts
…and 5 more, not listed.