AI is on both sides of the phishing fight now

AI now sits on both sides of the phishing fight. Attackers use it to write lures and hide their links, and Microsoft is about to use it inside Teams to read the context of a chat. Neither changes the basics, but both change where you should look.

What the attackers have now

Researchers at Abnormal Security have described a phishing-as-a-service platform called Athena StealthRoute. Its "AI Studio" has six workflows that produce tailored lures and decoy content, so a criminal no longer has to write convincing emails by hand. It also offers domain management and link cloaking, which sends security scanners to harmless pages while real targets get the phishing page.

Between April and 9 October, Abnormal says it saw 89,700 messages from the platform aimed at 2,280 organisations in the US and Europe. Note one detail: the platform does not steal passwords itself. It handles writing and delivery, and hands victims on to separate services that harvest credentials or tokens.

That split matters. If the email is well written and the link looks clean when your filter checks it, the old checks of spelling mistakes and bad links won't help much. The experts quoted suggest looking for shared back-end infrastructure and repeated link patterns across many messages, rather than judging each email alone.

What defenders are getting

Microsoft says Teams will soon use AI to detect phishing, impersonation and social engineering in chats. The key point is that it looks at the conversation, not just links and files. A message that says "I'm from IT, please approve the prompt on your phone" has no attachment to scan, but the intent is clear to a person, and Microsoft wants the system to see it too. It focuses on higher-risk cases such as a first message from someone outside the organisation.

Targeted release is due to start in early November 2026, with general availability expected from early to mid-December. Microsoft says no admin action is needed, and that message content is not kept for training or other uses.

What I'd actually do

None of this replaces the controls that already work. These are the ones I'd check first.

Tighten Teams external access. If you don't need open federation, allow only the domains you work with. Fake helpdesk chats from new external tenants are a common way in.
Use phishing-resistant sign-in where you can. If the platform hands victims on to credential and token harvesters, passkeys or FIDO2 keys for admins take away much of the prize.
Watch for token theft, not just password theft. Look at sign-in logs for odd locations and new devices straight after a user reports a phish, and revoke sessions rather than only resetting the password.
Teach the new tell. The tell-tale signs are fading. Train people to question the request (approve this, pay this, install this) rather than the spelling.
Report and pivot. When one user reports a phish, search the whole tenant for the same sender, subject or link pattern. That is the "look across messages" advice in practice.

The Teams feature is announced, not yet live. Treat it as an extra layer when it arrives, not a reason to relax external access settings now.

The bottom line

AI makes the attacker's email better and the defender's filter smarter, and they roughly cancel out. What still decides the outcome is the boring stuff: who can message your users, how they sign in, and how fast you react when someone reports something odd.

Sources

← All posts