Thanks again to Mohib for writing another great guest piece! This is another one of the important issues that we will be seeing more of in the coming years!
What are your thoughts Mohib on using a password manager as one of the tools to prevent a successful credential Phishing attack? If the web URL is even slightly wrong, the password manager won't fill in the login credentials, alerting the victim to the fact that something phishy is going on.
Obviously you need multiple layers of protection to keep yourself safe, but I believe that a password manager should be one of those layers.
Hey Chris, sorry for the late response. I think you’re right. Password managers can help protect you in this case if you’re using them through the browser extension.
But for people who don’t use that or are in the habit of opening the website themselves and copy-pasting their credentials, they could still fall victim to this kind of AI phishing attack.
Good point @Mohib Ur Rehman. I didn’t realise people used password managers without using the browser extension. That’s a vital bit of the puzzle.
I need to start posting myself about the importance of the password manager browser extensions (or equally linking the password manager to password autofill on your mobile / smartphone)
If your password manager could tell you that you were being phished that would be something. Heck if your email protection service just did its job, that would be something.
If you use your password managers browser extension to autofill login credentials then it will in effect tell you you are being Phished.
If the site is genuine the browser extension will fill in the username and password and login for you.
If it's a phishing site then the browser extension will not fill in your username and password, or log you in automatically, which alerts you that something is "Phishy" about the site.
The 54% and 12% match a study that is worth reading next to this post. Microsoft's one-page summary of the 2025 report says 54% for AI-automated phishing emails against 12% for standard attempts. The same numbers, and the same "AI-automated" wording, are in a paper by Heiding, Schneier and colleagues, first posted on arXiv in November 2024. It had 101 participants in four groups. A control group of arbitrary phishing emails got 12%. Emails written by human experts got 54%. Fully AI-automated emails got 54%. AI with a human in the loop got 56%.
So in that study the 12% is generic phishing, not hand-written spear phishing, and the experts writing by hand matched the AI exactly. What the automation changed is the cost per target. The paper puts the profitability gain at up to 50 times for large audiences.
That supports the post's scale argument more than the click rate does. The AI did not out-click a skilled human. It made the skilled human's result cheap enough to send to everyone.
The Arup employee did what the training says to do: got suspicious and asked for a video call. But the call was presumably set up through the same thread the attacker controlled, so the "second channel" was really the first channel with a camera attached. An independent channel has to be one the verifier opens on their own, from a directory the requester can't touch, like dialing the CFO's number from the internal phone list. The IBM figure (16 hours down to five minutes) points the same way. When a lure costs almost nothing, attackers can afford to fail nearly every time, so trying to win at the inbox is a bad trade. The control that holds up sits at the payment step, where one callback costs the defender a few minutes and costs the attacker the whole scheme.
Putting the control at the payment step makes sense to me, and I think it only works if the payment system enforces it. If the callback depends on an employee choosing to make it, the urgent requester can push them to skip it, which is the same pressure the original lure applies.
So payment release waits until the callback is recorded, ideally by a second person who didn't take the request.
Great article and collab, guys - on a really important topic. We'll see more deepfakes like this, and they'll keep getting more realistic, in companies and in families.
We'll keep pushing phishing-resistant MFA and the other controls wherever the systems allow it. But the human who has to make the call will still be the weakest point, so... we may also need to keep praying 🤣
Hehe, we'd pair the praying with the article's "boring fundamental" of independent verification, like calling the person back on a number you already have :)
The 54% vs 12% click rate stat is wild. One thing AI can't fix for the attacker: the link still has to point to a domain they registered, and that domain has a registrant, name servers and a history. Filters that only read the email text miss all of that.
True, but I think part of it is we're asking humans to judge stuff they can't even see. Nobody can tell a domain is 2 days old just by looking at the email. That check has to happen before the email reaches them. That's what we work on at WhoisFreaks (I'm on the team): flagging phishing domains through who registered them and what infrastructure they share, so the link gets blocked before anyone has to decide.
Thanks again to Mohib for writing another great guest piece! This is another one of the important issues that we will be seeing more of in the coming years!
Thanks for having me and I am loving doing this with you.
What are your thoughts Mohib on using a password manager as one of the tools to prevent a successful credential Phishing attack? If the web URL is even slightly wrong, the password manager won't fill in the login credentials, alerting the victim to the fact that something phishy is going on.
Obviously you need multiple layers of protection to keep yourself safe, but I believe that a password manager should be one of those layers.
i think Chris is right here. a great feature of using password managers. great point!
Hey Chris, sorry for the late response. I think you’re right. Password managers can help protect you in this case if you’re using them through the browser extension.
But for people who don’t use that or are in the habit of opening the website themselves and copy-pasting their credentials, they could still fall victim to this kind of AI phishing attack.
Good point @Mohib Ur Rehman. I didn’t realise people used password managers without using the browser extension. That’s a vital bit of the puzzle.
I need to start posting myself about the importance of the password manager browser extensions (or equally linking the password manager to password autofill on your mobile / smartphone)
If your password manager could tell you that you were being phished that would be something. Heck if your email protection service just did its job, that would be something.
True.
+1
If you use your password managers browser extension to autofill login credentials then it will in effect tell you you are being Phished.
If the site is genuine the browser extension will fill in the username and password and login for you.
If it's a phishing site then the browser extension will not fill in your username and password, or log you in automatically, which alerts you that something is "Phishy" about the site.
The 54% and 12% match a study that is worth reading next to this post. Microsoft's one-page summary of the 2025 report says 54% for AI-automated phishing emails against 12% for standard attempts. The same numbers, and the same "AI-automated" wording, are in a paper by Heiding, Schneier and colleagues, first posted on arXiv in November 2024. It had 101 participants in four groups. A control group of arbitrary phishing emails got 12%. Emails written by human experts got 54%. Fully AI-automated emails got 54%. AI with a human in the loop got 56%.
So in that study the 12% is generic phishing, not hand-written spear phishing, and the experts writing by hand matched the AI exactly. What the automation changed is the cost per target. The paper puts the profitability gain at up to 50 times for large audiences.
That supports the post's scale argument more than the click rate does. The AI did not out-click a skilled human. It made the skilled human's result cheap enough to send to everyone.
🔥🔥🔥 fantastic point and follow up.
Loved the post!
awesome thanks a ton. Really happy to be hosting Mohib and his expertise!
The Arup employee did what the training says to do: got suspicious and asked for a video call. But the call was presumably set up through the same thread the attacker controlled, so the "second channel" was really the first channel with a camera attached. An independent channel has to be one the verifier opens on their own, from a directory the requester can't touch, like dialing the CFO's number from the internal phone list. The IBM figure (16 hours down to five minutes) points the same way. When a lure costs almost nothing, attackers can afford to fail nearly every time, so trying to win at the inbox is a bad trade. The control that holds up sits at the payment step, where one callback costs the defender a few minutes and costs the attacker the whole scheme.
fantastic points and thanks for the comments!
Putting the control at the payment step makes sense to me, and I think it only works if the payment system enforces it. If the callback depends on an employee choosing to make it, the urgent requester can push them to skip it, which is the same pressure the original lure applies.
So payment release waits until the callback is recorded, ideally by a second person who didn't take the request.
100% great approache
Excellent (and scary) article!
really appreciate it Fabrice! really like how it landed. social engineering and phishing are really top threats!
Thanks for the read : )
Thank you for shedding light on this growing threat.
🔥🙏
Glad it was helpful to you!
Great article and collab, guys - on a really important topic. We'll see more deepfakes like this, and they'll keep getting more realistic, in companies and in families.
We'll keep pushing phishing-resistant MFA and the other controls wherever the systems allow it. But the human who has to make the call will still be the weakest point, so... we may also need to keep praying 🤣
Thanks for the kind words 🙏
thanks a ton Nelson! and i agree, it’s a growing capability for an attack that constantly ranks as a top threat!
Hehe, we'd pair the praying with the article's "boring fundamental" of independent verification, like calling the person back on a number you already have :)
100% the boring fundamentals still rock.
yessirr
The 54% vs 12% click rate stat is wild. One thing AI can't fix for the attacker: the link still has to point to a domain they registered, and that domain has a registrant, name servers and a history. Filters that only read the email text miss all of that.
totally agree. the click rate is crazy. and it reinforced what we’ve seen. humans are often the weak link.
True, but I think part of it is we're asking humans to judge stuff they can't even see. Nobody can tell a domain is 2 days old just by looking at the email. That check has to happen before the email reaches them. That's what we work on at WhoisFreaks (I'm on the team): flagging phishing domains through who registered them and what infrastructure they share, so the link gets blocked before anyone has to decide.
💯🔥🔥🔥