In April this year, Microsoft published a blog in which it asked a simple question: Do users really need third-party antivirus software or is Defender enough? The post said Microsoft’s built-in antivirus covers everyday risk, provided you’ve done all the necessary things, like installing updates, watching what you download, etc. And while they did say that it ultimately depends on what you get up to as a PC user, it was an inflammatory statement for the vendors who have been selling protection since the ‘90s.
The company removed the post in June after criticism from the cyber industry players, without any explanation. This isn’t to say Defender antivirus isn’t a great solution. It has to be – it’s built directly into Windows, which means it has nearly 1.6bn devices to keep safe. But Microsoft posed what many security analysts have been asking for years, which is whether a standalone antivirus is still necessary when so many of today’s attacks no longer involve what antivirus was built to catch in the first place: malware.
“The conversation has moved on from simply choosing a product,” says Angela Pringle, cybersecurity lead at CyberLogic. With offensive security, she says the biggest gap isn’t the technology, but what happens after it’s bought. Configuration, monitoring and alert response are what decides whether a control holds up against a real attacker, and endpoint protection now has to fit into a far wider strategy that includes identity, cloud, mobile and AI, which is a very long way from where antivirus started out.
Hand an AI agent a messy process and you just get a bad process happening faster.
Angela Pringle, CyberLogic
For decades, antivirus worked from a list. Vendors collected samples of known malware, wrote a signature for each one and pushed those signatures out to every machine they protected. It was a useful model built for a time when threats arrived as executables and spread slowly enough for the list to keep up. “That model has been outdated for years,” says Luis Corrons, security evangelist at Norton. “The first thing we need to do is stop thinking about antivirus as a product that simply scans files for malware.”
Modern endpoint protection looks at behaviour, processes, scripts, network activity, reputation and cloud intelligence before it reaches a verdict, and Corrons says the next stage is making that protection more contextual and more autonomous. “If a suspicious process starts accessing credentials, creating persistence and communicating with an unusual external service, the important question is no longer whether we have previously seen that particular file,” he says.
“It’s whether those actions make sense in that context and whether the system can stop the attack before significant damage is done.”
Being able to judge whether an action makes sense in context has become essential because so many of today’s attacks never touch a malicious file. CrowdStrike’s ‘2026 Global Threat Report’ found that up to 82% of detections last year were malware-free, with attackers working through stolen credentials, legitimate admin tools and trusted systems. Security teams call this “living off the land”, such as where an attacker uses whatever is already installed on a network rather than bringing in malware of their own.
“Signature-based antivirus was never designed for this, so relying on it alone is increasingly difficult to justify,” says Kurt Goodall, Troye technical director. A script that runs every day on a systems administrator’s machine looks very different on a laptop in the HR department, and Goodall says security has to account for the particular user, device, application and environment involved before it can tell the two apart. This also means questioning the assumption that buying more security products automatically makes an organisation safer, since the real value comes from how well those controls are integrated, monitored and managed. “A badly configured security stack is still a badly protected environment,” says Goodall.
In June, Sophos researchers stumbled on something unusual while investigating a compromised machine. A threat actor linked to ransomware and data theft had built a malware-testing lab in which several AI agents worked together to find ways of slipping past endpoint detection and response (EDR) tools, even going so far as to frame malicious prompts as red-team work.
AI, be it generative or agentic, has changed the economics of cybercrime. Reconnaissance, social engineering and scripting that once demanded real expertise can now be handed to a swarm of agents, which means attacks are happening quicker, and within reach of less skilled criminals. “They live off the land, abusing legitimate tools and stolen credentials, which gives traditional antivirus nothing obvious to flag,” says CyberLogic's Pringle, adding that keeping pace means building a security operation that moves at AI speed rather than looking for a product that “solves AI”. In theory, she says, antivirus can keep up, but only if it stops trying to work out whether something was written by a machine and focuses on what the attack does instead.
A compromised or poorly controlled AI agent could potentially become an attack path in its own right.
Kurt Goodall, Troye
“An AI-generated attack still has to do something,” says Norton’s Corrons. “Malware still needs to execute, steal credentials, manipulate processes, establish persistence, encrypt information or communicate with infrastructure controlled by the attacker.” Every one of those actions leaves a trace that security software can pick up on and security teams now have AI of their own to help spot it. Corrons says it’s more concerning that many AI-enabled attacks don’t begin with malware at all and, instead, go after a person’s identity, email account or decision-making process.
AI on the endpoint also changes what antivirus is being asked to protect. An agent can read information, use credentials, interact with several systems and take actions on an organisation’s behalf, all with permission. “A compromised or poorly controlled AI agent could potentially become an attack path in its own right,” says Goodall. He adds that the temptation will be to give agents broad access because it makes automation easier. A survey by Gravitee, an AI management platform, found that 54% of organisations had experienced, or suspected, an AI agent security or data privacy incident in 2026.
Corrons says an agent should only be allowed to make recommendations rather than decisions, and, once it’s performance is understood, it can then move on to limited, low-risk and preferably reversible actions. Anything more consequential should require a staff member’s approval. “It should operate with the minimum permissions necessary, its actions should be logged and auditable and there must be a straightforward way to revoke its access,” he says. Least privilege, logging and testing in a controlled environment are the same rules organisations already apply to privileged users and they only work if the process the agent inherits is sound in the first place. “Hand an AI agent a messy process and you just get a bad process happening faster,” says Pringle.
IT has traditionally looked after antivirus and the endpoints it runs on, while security teams increasingly look after detection, policy and incident response, and in many organisations, a managed service provider handles some, or all of it, which leaves plenty of room for responsibility to fall through the gaps. “ Saying that cybersecurity is everyone’s responsibility sounds good, but if taken literally, it can result in nobody being responsible for making sure a particular control works,” says Corrons.
End-users have a part to play, but an organisation whose protection depends on every employee correctly interpreting every security warning will eventually be let down by someone making the wrong call. “Buying an endpoint security licence is not the same thing as being secure,” says Goodall. This is one of the reasons Corrons says that smaller organisations rarely need an enormous EDR or extended detection and response (XDR) platform with dozens of capabilities that nobody has the time or expertise to run. Good endpoint protection, which is what antivirus has grown into, paired with a managed security service, will often do more than a sophisticated product left largely untouched. “The best endpoint security product is not the one with the longest feature list. It’s the one that is properly deployed, correctly configured, actively monitored and connected to a process that can respond when something goes wrong,” says Corrons.
* Article first published on www.brainstormmag.co.za
Share
