Learning Library

← Back to Library

AI vs Humans Crafting Phishing Emails

Key Points

  • Phishing attacks have become increasingly sophisticated, and a recent experiment compared the effectiveness of generative AI‑crafted phishing emails versus those written by humans.
  • IBM X‑Force researchers prompted generative AI to generate industry‑specific concerns, then instructed it to compose a socially engineered, marketing‑styled phishing email that leveraged empathy, FOMO, and urgent calls to action.
  • The AI‑generated email directly referenced the targeted concern (“limited advancement opportunities”), included personalized language, multimedia links, and a “click‑now” urgency to maximize the likelihood of victim compliance.
  • Human attackers countered by employing open‑source intelligence (OSINT) from sources like LinkedIn, Glassdoor, and corporate sites to gather detailed target information before crafting their own urgency‑driven phishing messages.
  • The study highlights that both AI and human tactics now use tailored, context‑aware content and time‑pressured prompts, underscoring the need for advanced awareness and detection strategies.

Full Transcript

# AI vs Humans Crafting Phishing Emails **Source:** [https://www.youtube.com/watch?v=7XhySwUn9eA](https://www.youtube.com/watch?v=7XhySwUn9eA) **Duration:** 00:10:55 ## Summary - Phishing attacks have become increasingly sophisticated, and a recent experiment compared the effectiveness of generative AI‑crafted phishing emails versus those written by humans. - IBM X‑Force researchers prompted generative AI to generate industry‑specific concerns, then instructed it to compose a socially engineered, marketing‑styled phishing email that leveraged empathy, FOMO, and urgent calls to action. - The AI‑generated email directly referenced the targeted concern (“limited advancement opportunities”), included personalized language, multimedia links, and a “click‑now” urgency to maximize the likelihood of victim compliance. - Human attackers countered by employing open‑source intelligence (OSINT) from sources like LinkedIn, Glassdoor, and corporate sites to gather detailed target information before crafting their own urgency‑driven phishing messages. - The study highlights that both AI and human tactics now use tailored, context‑aware content and time‑pressured prompts, underscoring the need for advanced awareness and detection strategies. ## Sections - [00:00:00](https://www.youtube.com/watch?v=7XhySwUn9eA&t=0s) **AI-Generated Phishing Email Test** - The speaker outlines IBM X‑Force’s experiment that used generative AI to create industry‑targeted phishing messages, comparing the AI‑crafted emails to those written by humans and highlighting the growing sophistication of AI‑assisted social engineering. ## Full Transcript
0:00you just got an email from a Nigerian 0:02prince he's got a ton of money and he's 0:04willing to share some of it with you all 0:07he needs is just your banking details 0:09and he can launder the money through 0:11your account please tell me you don't 0:13fall for that anymore that's a fishing 0:15attack and now fishing attacks have 0:16gotten a lot more sophisticated in fact 0:18they're about to get even more 0:20sophisticated and we ran an experiment 0:23where we looked at who's better at 0:25writing fishing emails these days humans 0:27or generative AI ding round one in this 0:31corner generative AI in this corner 0:35humans let's see how they did it well 0:37IBM's xforce researchers took a look at 0:40generative Ai and they asked it to come 0:42up with a list of concerns that people 0:45in a particular industry might have they 0:47targeted an industry so that the 0:49concerns would seem more relevant so it 0:52comes up with this list then they say 0:54okay what we want you to do is write an 0:57email leveraging social engineering 1:00techniques as well as marketing 1:03techniques and we're going to take all 1:05of this together and generate a fishing 1:07email using the generative Ai and 1:10finally they asked it who in fact should 1:13we send this to and who should it look 1:15like it came from they put all of that 1:18together and this is what came 1:21out Dear employees we understand that 1:23many of you are concerned about the 1:25issue of limited advancement 1:26opportunities see we got that from the 1:29list of concerns 1:30that were generated we want to make sure 1:33you have the resources you need to take 1:35your career to the next level this is 1:37all about you after all so it's included 1:39empathy and it's related to the person 1:42that's why I'm inviting you to a special 1:44event not just a normal event this is a 1:45special one to address the issue of 1:48limited advancement opportunities so 1:50it's hitting right on what your concerns 1:52are we understand that your time is 1:54precious so we're not going to waste it 1:57uh and we're going to even include with 1:59this a mob Global optimized website with 2:01stories and videos who doesn't love 2:03stories and videos with relevant 2:05information now at this point you've got 2:07to be having the feeling of missing out 2:09the fomo and we encourage you to act now 2:13there's got to be urgency do it now 2:15don't think just act to take advantage 2:17of this opportunity click on the 2:19link bang you've been fished dingding 2:23round two now the humans fight back 2:27let's see how we did well the approach 2:29for the exforce researchers that did 2:31this was they tried to leverage 2:33something called open-source 2:36intelligence that's looking at sources 2:38of information about people such as 2:40LinkedIn Glass Door Company websites and 2:43blogs and Gathering as much information 2:45as they could about the people they were 2:47going to Target that way they get the 2:49best information and can really Target 2:52the fishing email the next thing they 2:54did was create in their email that they 2:57crafted a sense of urgency we want 3:00people to act and not think if we're a 3:03Fisher so we're going to create some 3:05level of urgency some time constraint 3:07you need to act before Friday something 3:11like that and then the final element 3:13that they included was the notion of 3:16brevity we're not going to take a lot of 3:17your time we just got five simple 3:19questions this is a survey for instance 3:22or we want your opinion we want you to 3:25do the following things but it's not 3:26going to take much time so urgency along 3:29along with the fact that it's going to 3:31be brief now people are more likely to 3:33go ahead and do that so in this epic 3:35battle of man versus machine the winner 3:38is humans yay I don't know if this is a 3:42contest we necessarily wanted to win but 3:44we had a slight win it turned out that 3:47more people were fooled by the human 3:49generated fishing email than were fooled 3:52by the generative AI but it was very 3:55slight the difference but let's take a 3:57look at something else here another 3:59Factor to consider is that it took about 4:0216 hours for the human team to generate 4:05their winning fishing email why did it 4:08take so long well something like this 4:10this open- Source intelligence work that 4:12requires a lot of time to read through 4:14and research and try to pull out the 4:15exact right details and so forth and 4:18then to write in just the right things 4:20so that you include the levels of of 4:23urgency and simplicity uh that that just 4:27takes a while for a human to come up 4:28with all of that on the other side one 4:31person could go in to chat GPT and in 4:34five prompts in five minutes come out 4:37with something that was nearly as good 4:40so Effectiveness winner humans 4:43efficiency winner generative Ai and if 4:46you consider that this is continuing to 4:49improve we're going to improve a little 4:51bit but only so much this technolog is 4:54very new it's going to improve a lot so 4:56we're going to see improvements in this 4:58space where generative AI gets better 5:01and better at the way it does everything 5:03including the ability to write fishing 5:06emails now if you say but Jeff Chachi PT 5:08won't write fishing emails for me if I 5:10try to ask it to do that it's got guard 5:12rails good but there are prompt 5:14engineering prompt injection attacks 5:16where people get around those things 5:18also there are 5:21alternative to this type of generative 5:24AI there are alternative chatbots that 5:26have no guard rails and they will 5:28happily gener at all the fishing emails 5:30you want so we won sort of but we're 5:34going to lose in the long term unless we 5:36know what to do to deal with this threat 5:39okay so what can you do about these 5:41fishing attacks they're going to keep 5:42getting better and better what if we 5:44traditionally trained our users to do so 5:47that they don't fall for this well one 5:49of the main tells that we tell people to 5:52look for is bad 5:55grammar a lot of times the Fishers are 5:57not native English speakers and and 5:59they're writing in English and it looks 6:02like it so that can be a clue that uh 6:05someone could use to determine okay 6:07maybe this is a little suspicious 6:09especially if it seems to be claiming to 6:11come from an American company or a 6:14British company and yet the English is 6:16not very correct so that would be a good 6:18clue um another thing that we've often 6:21told people to look for is make sure 6:23that the thing is applicable if it's not 6:25applicable to you like I get an email 6:28from a bank that I don't do business 6:30with and they're asking for me to 6:31confirm my details then I know that's 6:34not for me so I can ignore it and then 6:37finally another major one that we tell 6:39people to look at look at the link look 6:41at the URL in the email that you're 6:43about to click on does it look bogus 6:45does it look like the normal link that 6:47you would use to go to your bank or to 6:49go to that particular shipping company 6:51or whatever it happens to be if it's not 6:54if it looks like there's a misspelling 6:56or it looks like it's in an odd format 6:58then we're going to ignore that so this 6:59has been the stuff that we've trained 7:01users to look for now how about with 7:04generative AI you saw the email that it 7:06came out with that was pretty good so 7:10one of the first things that we should 7:11be encouraging people to do is call use 7:15an out of- band communication to confirm 7:17that in fact this is a legitimate email 7:20and that this is a legitimate campaign 7:22for instance if the email says here's 7:24the phone number I'm going to ignore 7:26that what I'm going to do is say if 7:28you're my bank I know my Bank's phone 7:30number or I'm going to go look it up 7:32independently and I'm going to call the 7:34bank and I'm going to say did you send 7:35me this email should I click on this and 7:38if they confirm it okay that's a little 7:40bit different but this is one of our 7:42best defenses against fishing is an 7:44outof band confirmation like a call 7:47other things that we should unlearn is 7:51the thing that I just talked about was 7:52number one on the other list and that's 7:54looking for bad grammar you saw the 7:57fishing email the generative AI came up 7:59up with and the grammar in it was just 8:01fine so we have to tell users stop 8:04looking for that as a clue if they're 8:06looking for that as a clue and they 8:08don't see it then they will drop their 8:09defenses so in fact we have to ignore 8:12that one that has to be a new change 8:14another thing we need to do is expand 8:17the forms that we expect to see fishing 8:19coming in so for instance one type of 8:22this is called Vishing and in a Ving 8:26attack we're using voice so maybe we 8:29have a deep fake an imitation of 8:31someone's voice making a phone call to 8:33you and telling you to do certain things 8:35and you think you recognize the voice 8:37but you're actually not talking to that 8:39individual so we have to use the same 8:41kind of mindset the same kind of 8:43critical thinking and make a call back 8:46okay if this is really you I'm going to 8:48call you back at the well-known 8:50publicized number and see if you're 8:52still the same person that I can get to 8:53there another form of this is the SMS 8:57form of this smashing 9:00and in smishing attacks what we have is 9:03a text message that comes along in an 9:05SMS and this is going to contain a link 9:09with instructions and when I click on 9:11that going to end up with the same 9:13effect so in both of these cases it's 9:15the same type of attack it's just using 9:17a different Vector to expand that 9:20another thing that can really help here 9:22is a better use of identity and access 9:24management capabilities one of the 9:26things Fishers often do is try to steal 9:28your password how about if I don't have 9:30a password how about if I use a 9:33passwordless authentication capability 9:36using something like pass keys from the 9:39phto standard that this is something 9:41that I expect we'll see more of this 9:43grow no one can steal your password if 9:45you don't have a password in the first 9:46place and then I can make it even 9:48stronger if I use multiactor 9:50authentication not only something you 9:52know but something you are and something 9:55you have and combine all of these 9:57together and now when when someone tries 10:00to steal certain information about you 10:02they won't be able to have all of the 10:04other things like your actual face to 10:06unlock a system with or uh the 10:09particular phone that's been registered 10:10in advance so this makes it harder on 10:13the Fisher as well ultimately we have to 10:16keep 10:17adapting these fishing attacks are going 10:19to get more sophisticated and better 10:21over time remember the Nigerian prince 10:23we've come a long way from that point 10:26and it's going to only get better as 10:28generative ey gets better and does 10:30better types of attacks so one of my 10:32favorite sayings is if you're satisfied 10:34with your security so are the bad guys 10:37so never be satisfied always be on the 10:39defense always be on the 10:41lookout thanks for watching please 10:44remember to like this video And 10:45subscribe to this channel so we can 10:47continue to bring you content that 10:48matters to 10:52you