Trending Topics
Trending
AI agents will be just as susceptible to scams as humans
Are you on the agentic AI hype train? They promise lots in the way of supercharging productivity through autonomous decision-making capabilities, including mimicking human behaviour. But mimicking human behaviour may also become AI agents’ greatest weakness.
For all the exciting news coming out about AI agents from Google, Anthropic, OpenAI, or Salesforce, I don’t think most companies are ready to face the difficulty of adopting AI agents. The cybersecurity implications of AI agents are enormous, especially for the identity management market.
My prediction for 2025? When companies realise how much work security teams will have to do to retrofit existing security models to address AI agents, it’s going to slow AI deployments to a crawl.
AI agents have an identity crisis
The problem is that most solutions in the traditional identity management market have operated on a binary assumption: users are either human or machine. AI agents throw a monkey wrench into that assumption because they’re neither of these things. They straddle the line between machine and human.
You could technically call AI agents software, but they’re also unpredictable like humans, capable of independent actions like moving a mouse cursor. Mimic a human and you’re also mimicking their mistakes. And humans make a lot of mistakes.
This isn’t hypothetical. AI agents have already been found to procrastinate by browsing photos of Yellowstone if you give them enough freedom. Some of this comes down to AI agents being in early stages of maturity, but is any software immune to error? When was the last time you heard of a bug-free program? And since we’re talking about AI agents mimicking humans, when was the last time you heard of a human who never made a mistake?
One of the more common human mistakes that cybersecurity professionals have to deal with is falling for phishing scams. Who’s to say an AI agent won’t do the same? Not only is AI perfectly capable of lying to us, but one team of cybersecurity researchers even fooled a popular AI assistant into becoming a data pirate through indirect prompt injections.
The companies pouring money and research into AI are already keenly aware of how easily a user could teach an agent to “forget previous instructions”. OpenAI, for example, is training LLMs to prioritise privileged instructions. But if AI agents are made in the image of humans, they will also inherit human error, no matter how much we train them. Humans have already been trained (to varying degrees) to not click on suspicious links, yet identity attacks – and specifically password-based attacks – still make up a significant amount of cyberattacks.
As I’ve discussed in some of my previous articles, cybersecurity cannot coexist with human error, and AI is not some magic pill for safeguarding digital infrastructure. If AI agents are going to be successfully adopted and deployed, they have to be made immune to human mistakes. The clock is ticking, however, because 82% of the 1,100 executives surveyed by Capgemini want to implement AI agents in the next three years.
Download The Journey to Zero Trust
Embark on the journey to Zero Trust with this essential guide from Ping Identity. Discover how to implement a robust Zero Trust framework that enhances security, mitigates threats, and protects your organisation’s most valuable assets. Download this whitepaper today, to explore key strategies, best practices, and actionable steps to build a secure, Zero Trust architecture for your business.
The great consolidation of the identity management market
Ultimately, the more human-like AI agents become, the more important it will be to treat AI like humans. We’re not going to be countering software vulnerabilities as much as we’ll be countering human error. To do this efficiently, enterprises will have to adopt tools that don’t distinguish between machines and humans. Such a distinction will effectively disappear over time as AI agents continue to blur the line between software and humans.
I expect over time that the identity management market will consolidate, with more and more tools that offer unified or hybrid solutions for managing both human and machine identities at the same time. And that is a good thing, because identity fragmentation is a bad enough problem as it is, and the last thing the world needs is AI agents adding yet another identity silo for cybercriminals to exploit.
Beyond unifying identity management of humans and machines, another critical component to making AI agents immune to human mistakes is to end reliance on passwords and other static credentials. To illustrate just how widespread static credentials have become, one new report claims that over a billion passwords were stolen in 2024.
The first obvious measure all enterprises looking to adopt AI agents should take is to abolish the use of these credentials. Whether a user is AI or human, identity should never be presented as digital information stored on a computer. Cryptographic authentication has to be the way forward, and more than that, all access should only be temporary, and only allotted for specific tasks. The repeat-breaches of the Internet Archive showed us how easily malicious actors can repurpose exposed tokens from older incidents to infiltrate and move laterally across networks.
AI agents need zero trust, too
AI agents will undoubtedly impress and challenge us. However, their successful integration depends on organizations thoroughly addressing underlying identity and access management complexities. There’s no simple “on” switch for AI agent adoption.
We cannot eliminate mistakes – whether from humans or AI. But we can design infrastructure resilient enough to anticipate and mitigate potential vulnerabilities. If you were tired of hearing about zero trust, expect to see a lot more life injected into that conversation, because the same principles of least privilege and zero trust that applied to humans will also apply to humans.
The future of cybersecurity lies not in preventing errors, but in creating systems that can gracefully handle inevitable imperfections.
More from our Opinions section
About The Author
