Tech experts and business leaders have warned Metro that ‘shadow AI agents’ are posing an increasing security threat in the UK.
AI agents are systems designed to book travel, schedule meetings, sketch charts, handle customer complaints and even socialise with each other.
But these agents are being offered new gigs – as double-agents. Not in the 007 way, though.
Instead, as Microsoft tells Metro, ‘shadow AI’ are back-alleyway bots that have no formal approval or oversight from employers or officials.
A new poll of business leaders by Microsoft, shared exclusively today with Metro, shows that 84% consider shadow AI as a growing security threat.
What is a ‘shadow AI agent’?
AI agents need a lot of access and data to do their jobs – there are even guides these days on how to make them GDPR compliant.
These agents, Microsoft’s national security officer Jo Miller tells Metro, are common on personal and work phones and laptops.
‘We might choose to download some tools beyond Copilot, for example,’ she says of Microsoft’s AI model.
‘Some might be developed by Western companies, others elsewhere that have a different lens on how AI should be used and our data protected.
‘If I choose to download three more, maybe an image generator or a research agent, I can’t have the same confidence in where these tools come from – they could be harvesting my data and sharing it across the public internet, selling it, misusing it and playing it back as misinformation or disinformation.
‘There’s a range of risks that come with having AI tools on your device, on your network, and you don’t understand where they’ve come from and what they’re doing.’
What can shadow AI agents do?
Microsoft’s survey of 1,000 major public and private sector bosses, conducted in January, shows that bosses are quickly trying to get their heads around new-fangled tech like AI agents.
At least 62% of organisations are already deploying autonomous AI agents, almost tripling from 22% last year.
As much as shadowy AI agents are in the back of their minds, 68% expect agents to be fully integrated across their organisation within a year.
Microsoft says that as employees rush to embrace AI agents, they are creating security blind spots that bosses are addressing.
Most mainstream AI agents, Miller explains, have a level of autonomy held back by corporate guardrails – they won’t go off the rails, in other words.
AI agents used at work can sometimes be fully integrated – embedded in email services, slideshow software and other apps.
What do companies itching to use AI agents need to do to keep us safe?
Microsoft found that 86% of leaders are employing AI agents for security challenges. But 80% worry about managing agents at a large scale.
As the race is on to embrace these futuristic-sounding machines, 85% believe deployment is progressing faster than oversight approaches were built to support.
Nevertheless, 87% told Microsoft they’re confident they can prevent shifty AI tools from being created or used.
Security experts told Microsoft that they should have three priorities:
- Maintain visibility over where AI agents are operating (50%)
- Integrate agents safely into existing systems and processes (50%)
- Meet compliance, risk and audit requirements as autonomous activity expands (49%).
‘If I bring in another tool that will sit just outside our platform, I don’t know what back doors there might be to exfiltrate data,’ Miller says.
She adds: ‘We really need to be really deliberate and clear about what tools we’re downloading and using.
‘We don’t truly know where data might be going if we don’t understand the security parameters around a particular tool.’
This allows dodgy agentic tools to be used or exploited by cyber criminals or ‘hostile nation states’ to conduct cyber attacks, ransomware attacks, data theft and IP theft, actions sometimes described as ‘adversarial’.
By ‘hostile nation states’, also called nation-state threats, Miller means groups tied to countries with not the best intentions.
Think pro-Russia groups amid Moscow’s war against Ukraine, with Miller saying there has been a rise in cyber attacks over the last four years.
What should you do about shadow AI?
The main thing, according to Miller, is to only use AI tools you can trust.
‘Like by a known vendor or supplier,’ she adds, ‘that’s well-established and has published information around how secure they are.
‘There’s an element of faith or trust we place in AI, but we need to remember these tools are designed around the human brain.
‘So, in the same way a human brain misremembers, the same way the brain is not always factually correct, these models will not always be correct.
‘Humans in the loop adds a level of accountability and an assurance of output.’
Get in touch with our news team by emailing us at [email protected].
For more stories like this, check our news page.
MORE: Games Inbox: Are people right to criticise Nvidia’s DLSS 5 tech?
MORE: Val Kilmer to star in film a year after his death after being ‘brought back to life’
MORE: Capcom developers didn’t know about Nvidia DLSS 5 before reveal claims report
