AI agents are becoming identities in their own right, and security needs to catch up

The next phase of artificial intelligence is likely to create an unusual identity problem for businesses. Increasingly, the entity attempting to access an online account, purchase a product, change a booking or interrogate a corporate database may not be a person. It may be an AI agent acting for them.
This transition from generative to “agentic” AI is important because agents do more than produce content. They can be given objectives, determine intermediate steps and, where permissions allow, perform actions across multiple services. Enterprise products are already moving in this direction, with agents increasingly able to interact with tickets, repositories, messages and other business systems. Recent developments covered by Digital Journal include systems that can identify outstanding work, prepare the next action and execute approved changes.
This changes the problem from securing AI to securing what AI can do.
One implication is that businesses may need to distinguish between the customer and the digital agent authorised to represent that customer. Recent Digital Journal analysis of AI and digital identity considered precisely this problem: when an agent books a flight, makes a purchase or changes an insurance policy, the organisation receiving the instruction needs to know not simply who the customer is, but what software is acting for them and what authority has been delegated to it.
Dan Moore, Sr. Director of CIAM Strategy & Identity Standards at FusionAuth, expects this to become highly visible to consumers. “Major retailers like Target will have a link on their website asking folks to ‘add Target to your agent’ in the same way they prompt for you to allow notifications or install an app,” he predicts.
The comparison is useful. Instead of quietly sharing a username and password with autonomous software, consumers could explicitly establish a service relationship with an agent. That creates the possibility of granting the software narrowly defined permissions which can subsequently be monitored or withdrawn.
This is increasingly important because allowing an agent to inherit ordinary human credentials can make it difficult for a service to determine whether an action was initiated by a person or software. Moore predicts there will be a high-profile incident in which “an agent will cause someone to lose a significant amount of money or reputation due to using human credentials directly”.
Machine speed changes the risk
Erin Hamm, Sr. Director and Field Chief Data Officer at DataBee, A Comcast Company, sees the issue becoming harder as agents penetrate corporate systems. “In 2027, organizations will have to rethink identity and access management for a world in which employees are no longer the only entities acting with their permissions,” Hamm tells Digital Journal.
“When an AI agent operates on someone’s behalf, it is effectively borrowing that person’s identity. It can use that employee’s credentials and privileges to take actions at machine speed, potentially across multiple systems.”
This extends a familiar cybersecurity principle: least privilege. A person should only possess the permissions necessary to perform their role. Agentic AI makes failures in this model potentially more consequential because software can operate faster and across more systems than its human owner. As Hamm puts it: “An overprivileged employee with an AI agent is no longer simply one overprivileged identity.”
This requires organisations to think about agents as a new type of non-human identity. Controls could include separate credentials, restricted scopes, short-lived access tokens, detailed audit trails and explicit approval barriers before high-consequence actions. The principle is consistent with broader developments in cybersecurity, where identity is increasingly regarded as a security perimeter. Previous Digital Journal material has highlighted the need for non-human identities, including AI agents and system integrations, to be subjected to appropriate lifecycle management and scrutiny.
AI is finding the vulnerabilities humans missed
AI is simultaneously transforming another part of cybersecurity: vulnerability discovery. Scott Miserendino, Chief Technology Officer at DataBee, argues that frontier AI is providing an unprecedented ability to examine existing software.
“Frontier AI has dramatically increased our ability to find vulnerabilities, including bugs buried in software that has never been subjected to this level of computationally intensive, automated review,” he tells Digital Journal.
In one sense, AI is exposing what Miserendino describes as “years of accumulated software development debt”. This effect is already visible. Digital Journal has reported on AI-assisted discoveries of very serious software vulnerabilities, illustrating how advanced models can examine code at a scale that would be difficult to reproduce through human review alone.
Yet Miserendino does not expect the resulting surge in vulnerability discoveries to continue indefinitely. “There’s a finite backlog of existing software to scrutinize,” he explains. As organisations analyse more historic code and remedy discovered defects, this backlog should gradually diminish.
There is also a second effect. AI does not have to wait until code has entered production before finding faults. Models can increasingly analyse software during development. “The same capabilities finding vulnerabilities in existing software can increasingly review code and help developers identify and fix bugs before that software ever reaches production,” Miserendino says.
The result could ultimately be a counterintuitive security trend: AI initially increases the number of reported vulnerabilities because it becomes better at discovering them, before helping to reduce vulnerabilities by shifting detection earlier into software development. Previous Digital Journal reporting on AI-assisted vulnerability discovery similarly found that AI is changing the economics of identifying weaknesses, particularly through its ability to revisit legacy code at scale.
Amid these technological changes, however, Moore makes a less futuristic prediction: “Passwords will still be the way most people authenticate with most applications.” That is a useful reminder that technological transitions rarely happen cleanly. Passkeys, multifactor authentication, password managers and stronger password hashing can improve security, but legacy authentication systems remain widespread. Research discussed in New research shows, for example, highlights the importance of combining modern password hashing with controls such as breached-password screening and multifactor authentication because no hashing algorithm protects a password already compromised through phishing, malware or reuse. Consequently, enterprises face the uncomfortable prospect of managing highly autonomous AI agents alongside decades-old authentication architecture.
Then comes the quantum question
A further identity transition is appearing on the horizon. Moore expects “identity vendors will start shipping quantum-resistant signing and token options as auditors and governments begin asking uncomfortable questions.”
This is no longer entirely speculative. Post-quantum digital signatures are starting to appear in commercial security applications. Recent implementations have incorporated ML-DSA, the module-lattice-based digital signature algorithm standardised under NIST FIPS 204, for credentials intended to remain resistant to future quantum attacks.
This does not mean cryptographically relevant quantum computers have arrived. It means organisations responsible for long-lived identity infrastructure are beginning to consider the migration path before they do. Taken together, the developments point towards a different model of cybersecurity. The question will no longer simply be“Is this person who they claim to be?”
Security systems will increasingly need to establish which human authorised which agent, what that agent is permitted to do, what credentials it possesses, where it has operated and whether its actions remain within the authority originally granted. In the agentic era, establishing identity may therefore become less about authenticating a single user and more about governing a chain of delegated trust.
What does this really mean?
Sign up for our weekly newsletter where we look at the technology stories reshaping how Canadian businesses operate and compete, and what they mean for the people running them.
We land in your inbox on Tuesdays. You can unsubscribe at any time.

Dr. Tim Sandle is Digital Journal’s Editor-at-Large for science news. Tim specializes in science, technology, environmental, business, and health journalism. He is additionally a practising microbiologist; and an author. He is also interested in history, politics and current affairs.
The latest Artificial Intelligence news:
- Musk, Uber co-founder join call for Jensen Huang’s return to Nvidia November 16, 2026 5:04 AM EST

Musk, Uber co-founder join call for Jensen Huang’s return to Nvidia
- From Carnegie to Claude: AI wealth could reshape philanthropy November 16, 2026 5:01 AM EST

From Carnegie to Claude: AI wealth could reshape philanthropy
- Five things to know about Anthropic November 16, 2026 4:47 AM EST

Five things to know about Anthropic
- AI agents are becoming identities in their own right, and security needs to catch up November 15, 2026 11:39 PM EST
AI agents are becoming identities in their own right, and security needs to catch up
- Italy fines Meta $880 mn over photo privacy: authority November 14, 2026

Italy fines Meta $880 mn over photo privacy: authority
Related News
- An AI bug hunt leaves months of work for the security team Microsoft’s Linux research shows how much checking and repair can follow an automated search for software flaws. By Jennifer Friesen November 13, 2026

An AI bug hunt leaves months of work for the security team
Microsoft’s Linux research shows how much checking and repair can follow an automated search for software flaws.
- PHP2Shell exposes cybersecurity’s new AI paradox: Finding flaws is becoming easier than fixing them Artificial intelligence can uncover serious vulnerabilities at unprecedented speed. For the UK technology channel, however, PHP2Shell suggests the harder problem is increasingly what happens after a flaw has been found. By Dr. Tim Sandle November 12, 2026

PHP2Shell exposes cybersecurity’s new AI paradox: Finding flaws is becoming easier than fixing them
Artificial intelligence can uncover serious vulnerabilities at unprecedented speed. For the UK technology channel, however, PHP2Shell suggests the harder problem is increasingly what happens after a flaw has been found.
- AI is accelerating cyberattacks: Why Canada needs to prepare for a new era of digital risk Artificial intelligence is making cyberattacks faster, cheaper and easier to scale. For Canadian organisations, the challenge is increasingly shifting from attempting to prevent every incident towards building systems capable of surviving and recovering from attacks. By Dr. Tim Sandle November 11, 2026
AI is accelerating cyberattacks: Why Canada needs to prepare for a new era of digital risk
Artificial intelligence is making cyberattacks faster, cheaper and easier to scale. For Canadian organisations, the challenge is increasingly shifting from attempting to prevent every incident towards building systems capable of surviving and recovering from attacks.
- AI is forcing business owners to confront a bigger problem: How their companies actually work Artificial intelligence promises efficiency, but its deeper impact may be diagnostic, exposing businesses that depend too heavily on individual employees, undocumented processes and concentrated customer relationships. By Dr. Tim Sandle November 10, 2026

AI is forcing business owners to confront a bigger problem: How their companies actually work
Artificial intelligence promises efficiency, but its deeper impact may be diagnostic, exposing businesses that depend too heavily on individual employees, undocumented processes and concentrated customer relationships.
- AI could make writing code the easy part: Five ways enterprise software will change by 2027 As artificial intelligence makes software quicker to produce, the harder problem for businesses may become keeping AI-rich systems secure, reliable and economically useful. Simform executives predict this will reshape engineering budgets, development practices and the way autonomous agents are governed. By Dr. Tim Sandle November 9, 2026

AI could make writing code the easy part: Five ways enterprise software will change by 2027
As artificial intelligence makes software quicker to produce, the harder problem for businesses may become keeping AI-rich systems secure, reliable and economically useful. Simform executives predict this will reshape engineering budgets, development practices and the way autonomous agents are governed.
- Canada finally gets real-time payments, but the bigger banking transformation is still to come Canada’s long-delayed Real-Time Rail brings instant settlement to the country’s financial system. Yet without fully operational consumer-driven banking, the technology’s potential for competition, innovation and consumer control will remain only partly realised. By Dr. Tim Sandle November 9, 2026

Canada finally gets real-time payments, but the bigger banking transformation is still to come
Canada’s long-delayed Real-Time Rail brings instant settlement to the country’s financial system. Yet without fully operational consumer-driven banking, the technology’s potential for competition, innovation and consumer control will remain only partly realised.
- Still using ‘123456’? Why weak passwords remain an open door for cybercriminals High-profile security incidents are putting digital security back under scrutiny. Yet for millions of internet users, one of the weakest points in their defences remains something remarkably basic: the password. By Dr. Tim Sandle November 8, 2026

Still using ‘123456’? Why weak passwords remain an open door for cybercriminals
High-profile security incidents are putting digital security back under scrutiny. Yet for millions of internet users, one of the weakest points in their defences remains something remarkably basic: the password.