AI is most useful on defense when it takes the heavy, repetitive work off human hands and gives the judgment back to a person. In practice, that means using it to accelerate investigation: pulling context together and triaging high-volume signals so the clear-cut cases get handled fast. That speed matters because attackers now rotate infrastructure and lures quickly, and isolated alerts pile up faster than a lean team can work through them.

The part that keeps AI honest is human ownership. Analysts and threat hunters still own the complex investigations and the response decisions, because those calls carry consequences a model shouldn’t make on its own. The aim is to match machine speed with human judgment: let AI handle the legwork so people spend their attention where it counts. That pairing is how a lean team defends at the pace of the attack without standing up a 24/7 operation of its own.

The most common AI-powered threats are the familiar ones running at a new speed and scale. AI-accelerated phishing leads the list: tailored lures sent across email, text, calendar invites, and collaboration tools, written to match the target’s actual job. Closely tied to it is identity abuse, where attackers use stolen credentials, session cookies, device codes, or OAuth tokens to operate through a real account after a successful login. Business email compromise fits here too, since a compromised inbox lets an attacker study real conversations and insert fraud into a process people already trust.

Device code phishing has become one of the clearest examples. It abuses a legitimate login step, needs no malware or stolen password, and just asks a victim to do something that feels routine, which makes it cheap to run at volume and hard to flag. Huntress observed device code phishing attacks climb 1,380% between July–December 2025 and January–April 2026. On the endpoint side, human-enabled compromise (fake app installs, ClickFix-style “run this to fix it” prompts) rounds out the pattern, all of it built to pass for a routine task rather than obvious malware.

Defending against AI-enabled attacks starts with a shift in what you watch for. Because these attacks are built to look legitimate, the defense is coordinated visibility across identity, users, and endpoints, so a signal that looks fine in one layer gets caught when it doesn’t line up with the others. A login can succeed and still be wrong if it comes from the wrong device or a location that’s out of step with someone’s role.

A few habits do most of the work. Behavioral detection catches what signature-based tools miss, which matters against AI-generated tools that have never existed before and won’t match a known pattern. Training built around real tactics helps people question urgency and unexpected requests, even when those requests arrive through channels they already trust. And making reporting fast and blame-free means that when a convincing lure does work, defenders hear about it early enough to contain the damage. The thread running through all of it: question the workflow, not just the message.

Huntress treats phishing as a full attack path, because the message is usually only the setup. It starts at the people layer: Managed Security Awareness Training exposes users to the tactics attackers are actually using now, so recognition and reporting improve against current tradecraft rather than last year’s tells. When a lure does succeed, the important signals show up after the login.

That’s where Managed ITDR comes in, watching Microsoft 365 and identity behavior for the things that follow a successful phish: suspicious authentication, session and token abuse, rogue OAuth apps, and mailbox rule changes. If the attack pushes to the endpoint instead, through a fake app or a “run this fix” prompt, Managed EDR catches the infostealers and remote-access tools that follow. Tying it together is the 24/7 SOC, where AI accelerates the investigation and human analysts own the verdict. That combination is how Huntress catches an attack that looks like normal work, and it’s the reason coordinated visibility beats any single filter.

AI is changing social engineering mostly by removing its old limits. Social engineering always relied on human effort, since a person could only research so many targets and write so many convincing emails in a day. AI erases that ceiling. The same tailored lure that once took real time to craft can now be produced at volume and personalized from public information, then sent to far more people with far less manual work.

It’s also making the lures better. AI helps attackers polish impersonations and tune a message to a specific job or moment, mimicking someone’s writing voice closely enough to remove many of the signs people were taught to spot. In one campaign Huntress investigated, attackers used AI to scan a compromised inbox and calendar, then wrote a wire fraud email in the victim’s own voice within minutes. The manipulation tactics are the same ones that have always worked. What’s new is how cheap and convincing they’ve become at scale, which is why the real skill now is noticing when something perfectly ordinary is being used for the wrong purpose.