49 min
Fraudologyへようこそ。
シカゴで開催されたMerchant Fraud Allianceカンファレンスの開会にあたり、私が行ったプレゼンテーションをご紹介します。これは、不正攻撃を取り巻く現状を俯瞰する内容で、カンファレンスを通じて加盟店の皆さんから何度も耳にした疑問をきっかけに生まれました。現在の不正は、これまでと実際に何が違うのでしょうか。そして、その違いのうち、どれほどがAIによるものなのでしょうか？
Eコマースの不正対策に携わっている方なら、自社の数値からも変化を感じていることでしょう。生成AIによって、攻撃を仕掛けられる人の範囲や攻撃のスピード、さらには優良顧客に紛れ込む巧妙さまで変わりました。本プレゼンテーションでは、多くの加盟店との対話から得た俯瞰的な視点でAIを悪用した不正の実態を解説し、自社のチームや経営陣に持ち帰って活用できる不正防止策も併せて紹介します。
このエピソードでお届けする内容：
- 不正行為が産業化した理由と、AIによって大規模なEC攻撃に必要なコスト、スキル、時間がどのように削減されるのか
- 主な数値として、不正圧力は前年比33%増、アカウント乗っ取りは78%増、カードテスティングは175%増、BOPIS（オンライン購入・店舗受け取り）における不正は65%増
- 2015年のEMVチップへの移行やアカウント乗っ取りの増加と、現在AIが不正攻撃の状況にもたらしている変化との比較
- AIを悪用したアカウント乗っ取りが増え続ける理由と、生成AIによってフィッシングの個別最適化が低コストで可能になる仕組み
- AIを使ったカードテスティング攻撃は従来のスクリプトとどう異なるのか――AIエージェントがカードを選別し、対策に適応する仕組みも含めて
- 1つの攻撃キャンペーンが本人確認から決済、カスタマーサポート、異議申し立てへと移行していく仕組みと、イベント単位の対策ではそれを見逃してしまう理由
- 偽造レシートや偽造身分証明書から、偽の損傷写真やディープフェイクに至るまで、証拠を低コストで捏造できるようになった理由
- ダークウェブ上のカード売買市場のデモから見えてくる、攻撃に必要な要素をいかに簡単に購入できるか
- セッションレベルのデータ、テレメトリ、デバイスID、行動バイオメトリクスを活用したAI不正検知で、AIによる不正に対抗する方法
次のような方は、ぜひこのエピソードをお聴きください：
- 不正行為の増加について経営陣に説明する必要があり、その際に提示できる最新の数値が欲しい
- アカウント乗っ取りやカードテスティングが急増している、または決済代行会社からVisaの列挙攻撃対策プログラムについて連絡を受けた
- 写真や領収書、商品未着の申し立てが関係する返金詐欺やチャージバックに対処している
- 不正防止ツールを評価しており、スナップショットベースの対策ではなぜ攻撃を見逃し続けるのか疑問に思っている
- 加盟店が生成AIを活用して、AIが作成した詐欺をどのように見抜いているのか知りたい
- Merchant Fraud Allianceに参加できず、カンファレンスがどのように幕を開けたのか知りたい方
エピソードノート
詐欺における過去2回の転換と、AIが第3の転換となる理由
最初の変化は、EMVチップ搭載カードへの移行でした。チップが導入される以前、米国のカードには偽造しやすい磁気ストライプが使われており、不正利用は対面取引とオンライン取引でほぼ半々でした。英国などの他市場では、チップ導入後に何が起きるかがすでに明らかになっていました。対面取引での不正は減少し、オンラインでの不正は増加するのです。私は2013年から加盟店にこの点を警告し始めました。そして2015年10月にライアビリティシフトが実施されると、それらの市場での予測どおり、オンライン不正が増加しました。
その後まもなく、第二の変化が訪れました。加盟店がカード不正を抑え込むようになると、犯罪者は顧客アカウントに狙いを移しました。アカウント乗っ取りへの対策が必要になったことで、不正との戦いの場は決済ページだけでなく、ログインページにも広がりました。
AIは第3の転換点です。最初の転換は、他の市場ですでに起きていたため、予見できました。
数字で見る、組織化・大規模化する不正
主要な数字は無視できないほど深刻です。不正行為の圧力は前年比で推定33％増加し、アカウント乗っ取りは78％、カードテスティングは175％、オンラインで購入して店舗で受け取る（BOPIS）仕組みを悪用した不正は65％増加しています。インターポールの推計によると、AIを活用した手口は従来型の攻撃の約4.5倍の利益をもたらし、現在では小売業を狙った不正行為の約10件に3件がAIによって生成されているとみられます。この割合は来年、さらに高まるでしょう。私たちの目標は、これまでも不正を完全になくすことではありません。狙われるなら他のサイトよりも攻撃しにくく、時間とコストがかかるサイトにすることです。これは、熊から逃げるときは一番足の遅い人より速く走ればよい、という昔ながらの考え方です。しかしAIは、攻撃をより迅速かつ安価で容易に実行できるようにし、こうした優位性をことごとく縮小させています。
AIを活用したアカウント乗っ取り攻撃とカードテスティング攻撃
スクリプトは同じ動作を何度も繰り返すため、検知できます。一方、AIエージェントは状況に適応します。サイトで何が許可され、何が許可されないかを把握し、それに応じて動作を変えます。これが、アカウント乗っ取りが増え続けている大きな理由の一つです。生成AIを使えばフィッシングを簡単に個人向けにカスタマイズできます。また、多くのメッセージは、たいていの人がすでにアカウントを持っている配送会社、大手小売業者、決済ブランドなどを装うため、ごく普通の連絡に見えます。カードテスティングも同じ仕組みです。何千件ものカードデータをAIツールに取り込み、BIN、承認結果、住所ごとにカードを分類し、少額決済でテストできます。攻撃者は、まず別のサイトでテストすることもできます。そのため、それらのカードが自社サイトで使われる段階では、大量の決済拒否が発生しない可能性があります。これは、私たちの多くが長年頼りにしてきたパターンを崩すものです。
顧客ライフサイクル全体にわたる一つのキャンペーン
かつて、不正行為は複数の専門家によって分業されていました。あるグループは本人確認情報を扱い、別のグループはカードテスティングを行い、さらに別のグループは現金化を担当し、顧客サービスへの対応やチャージバックの申請を専門とするグループもいました。また、「Fraud as a Service（サービスとしての不正）」を通じて、互いにサービスを売買することもありました。しかしAIエージェントを使えば、1つのキャンペーンですべてを実行できます。盗まれた、または合成された身元情報を使って商品を購入し、サポートに連絡して商品が届かなかったと主張したうえで、完璧に整った証拠書類を添えて異議申し立てを行えます。その証拠書類自体も偽造されているからです。アカウント乗っ取りは、多くの場合、ロイヤルティポイントやチャージ残高の使い切り、あるいはマーケットプレイス出品者のアカウントに登録された支払先の変更という形で終わります。イベント単位の管理では、こうした不正を見逃してしまいます。承認、ログイン、本人確認情報を個別に見ると、いずれも正当に見えるためです。損失が明らかになるのは、多数の取引が正常に完了した後です。つまり、攻撃が終わるまで、攻撃を受けていることに気づかない可能性があります。
安価な証拠、安価なオンライン店舗
偽の領収書、注文情報、破損写真、異議申し立ての説明文、音声通話、本人確認書類は、今やすべて必要に応じて作成できます。ある大手フードデリバリー企業は、MFAの詳細調査で、破損写真が返金制度の不正利用を引き起こす最大の要因の一つだと述べています。チャットボットに頼めば、写真にカビを加えたり、1ダースの卵のうち5個を割れた状態にしたりできます。小売業者も、衣類に穴を加えた同様の事例を目にしています。また、現在ではディープフェイクの検出はほぼ不可能に近く、ディープフェイクを使った本人確認詐欺によって本人確認の安全性が脅かされていると考えています。偽のオンラインストアも同じくらい安価に作れます。Matt Vegaが実演したように、50ドル未満、わずか10分ほどで販売業者のサイトそっくりのコピーを作り、1文字だけ異なるURLを使ってソーシャルメディア広告から誘導できます。これは、広告詐欺とフィッシングが連携した手口です。今年発生したある攻撃では、オンラインのカードスキマーによって100以上の小売サイトが一斉に侵害され、60万件を超えるカード情報が盗まれました。また、Mattが録画したダークウェブ上のカード売買市場のデモも紹介しました。そこでは、購入者が国、州、郵便番号、カードブランド、BINで絞り込めます。AIが新たな手口を生み出しているわけではありません。従来の手口を正規の顧客に紛れ込ませやすくしているのです。
セッション全体を読み解き、詐欺師一人ひとりの兆候を見抜く
AIを悪用した攻撃を検知するチームは、顧客が商品を見つけてから購入後に至るまで、アイデンティティ、意図、コンテキストを結び付けて分析します。個々のイベントだけでなく、セッションをまたいだパターンを読み取るのです。私はよく防犯カメラにたとえて説明します。ログイン時や決済時のスナップショットは、誰かが家に侵入している瞬間を捉えた静止画のようなものです。侵入されたことは分かっても、どうやって入ったのか、何を盗まれたのかまでは分かりません。不正検知におけるセッションレベルのデータは、家の内外に設置されたカメラのように機能します。私が司会を務めたセッションで、Soups Ranjanは、どの詐欺師にも特有の兆候があり、テレメトリを使えば、スマートフォンを不自然な持ち方で操作している、あるいは画面を下にして置いているといった兆候を見つけられると述べました。私が見てきた中で、チャージバックと誤検知を低く抑えている先進的な企業は、デバイスIDと行動バイオメトリクスを組み合わせ、適切な不正シグナルを取り込んでいます。たった1つのシグナルを見逃すだけでも、損失やチャージバック、誤検知につながりかねないからです。
経営陣への説明と、創意工夫を凝らした取り組み
経営陣に急増の理由を説明する必要があるなら、業界のデータを活用しましょう。私の知る複数の加盟店では、春以降、アカウント乗っ取りやカードテスティングが突然増えた理由を説明しなければならなくなっており、決済代行会社からVisaの列挙攻撃対策プログラムについて連絡を受けたところもあります。Merchant Risk Councilの報告によると、加盟店の80%が不正対策における最大の課題としてテクノロジー基盤を挙げています。つまり、本当のハードルは適切なツールの導入承認を得ることだと私は考えています。カスタマージャーニーのデータは、その助けになります。マーケティング部門は顧客がサイト上で何をしているかを把握したがるため、予算を申請する際の味方になってくれるからです。一方、経営陣はより少ないリソースでより多くの成果を出すようチームに求めており、工夫を凝らして対応している加盟店もあります。あるマーケットプレイスのチームは、実際の買い手と売り手のチャット、およびフィッシングが疑われるチャットを生成AIツールに読み込ませ、AIが作成したメッセージに共通する特徴を尋ねました。その結果、冒頭のあいさつ、自由回答形式の質問の連続、そして最後にリンクのクリックを促す、というパターンが見つかりました。AIはAIを見抜けることが多く、現在では、手動審査の案件を精査して新たなルールを提案するAIエージェントを提供するベンダーもあります。AIにはAIで対抗できます。そして、効果のあった方法を共有することが、変化に遅れず対応していくための鍵です。
重要なポイント
- AIによってEコマースを攻撃するために必要なコスト、スキル、時間が削減されたため、私は不正行為が産業化していると表現しています。
- 不正行為の増加圧力、アカウント乗っ取り、カードテスティング、BOPIS（オンライン購入・店舗受け取り）を悪用した不正はいずれも急増しており、その主な要因の一つがAIです。
- インターポールは、AIを活用した詐欺の収益が従来型の攻撃の約4.5倍に上り、小売業界における不正行為の試みの約10件に3件が、すでにAIによって生成されている可能性があると推定しています。
- スクリプトは同じ動作を繰り返す一方、AIは状況に応じて適応するため、AIを用いたカードテスティング攻撃やアカウント乗っ取りは、従来のルールでは検知が難しくなります。
- 単一のAI主導型キャンペーンが、本人確認、決済、カスタマーサポート、異議申し立てにまたがることもあります。個々のイベントしか監視しない対策では、こうした動きを見逃してしまいます。
- 偽の領収書や身分証明書、破損写真は低コストで作成できるため、証拠だけを基に返金制度の悪用やチャージバックの妥当性を判断することが難しくなっています。
- ディープフェイクを目視で見破ることはほぼ不可能です。本人確認には、行動シグナルやデバイスシグナルによる補完が必要です。
- セッションレベルのデータ、テレメトリ、デバイスID、行動バイオメトリクスを活用することで、ある時点のスナップショットだけでなく、ユーザーの行動全体を把握できます。
- 業界の数値データがあれば、増加の理由を経営陣に説得力をもって説明し、より優れたツールの導入を求めることができます。
- 加盟店やベンダーはすでにAIを活用してAIによる攻撃に対抗しており、効果的な手法を共有することで、どのチームも攻撃を受けにくくなります。
最後に押さえておきたいポイント
このエピソードから一つだけ持ち帰ってほしいことがあるとすれば、私たちが対峙している不正は決して新しいものではない、ということです。従来からある不正が、より速く、より低コストで、より大規模に行われているのです。だからこそ、次々と現れる新たな手口を一つひとつ追いかけるのが答えではありません。顧客体験の流れをより広く捉え、適切なシグナルを取り込み、具体的な数値を示して経営陣を巻き込むことが重要です。そうすることで、導入しているツールも変化のスピードについていけるようになります。不正率の上昇は、単に業界全体と同じ傾向にあることを意味しているだけかもしれません。そこで次に何をするかが、他社との差を生みます。
エピソードの参考資料とリンク
つながる：Karisse Hendrick | LinkedIn
Fraudologyポッドキャストのホスト
受賞歴のあるサイバー詐欺対策の専門家
Eコマース不正防止コンサルタント
スタートアップアドバイザー、基調講演者、
Fortune 500企業向けコンサルタント
Episode transcript
Karisse Hendrick
00:07
Welcome back to the Fraudology podcast. I'm Karisse Hendrick. Well, we had a guest need to reschedule this week, so you get another solo episode with me. Last week, we talked about the Merchant Fraud Alliance and how just amazing the first uh conference was. Uh we are already talking about next year. Um I will give you more details on that when I have them. But I decided that maybe what I would do today is share a bit or share all of the uh presentation that I did to kick off MFA. Um I didn't want to call it a keynote because I didn't think that's what it was. It was more of just like a here's where we are and fraud presentation. Um talking about the fraud landscape and you can't talk about the fraud landscape right now without talking about AI and specifically how AI is attacking and changing fraud. It reminds me of back in 2015 uh actually even before that starting in 2013 I started to talk a lot about EMV in my presentations. I worked for the trade association in our industry at the time. I did a lot of presentations at other conferences as well as our own and uh I was talking a lot about EMV and how that was going to change online fraud forever. Um essentially in the US we were using cards with the magnetic stripe which was very easy to counterfeit back then. So the breakdown of fraud between card present fraud and card not present fraud was about 50/50. But we had seen in the UK as well as in Africa and other markets that when the banks changed their cards to include an EMV chip uh chip and pin or just chip that online fraud went up significantly and in-person fraud went down quite a bit. And you know there's cons some conspiracy theorists that were saying hm isn't that interesting that banks are doing something to reduce in-person fraud because per the credit card rules issuing banks take the hit they take the liability on card present transactions. But because the chips made it much more difficult to counterfeit cards, we saw we knew we were going to see online fraud go up. And that's exactly what happened in October of 2015 when that law changed the rule the law that required banks to include chips on their cards. So I went around preaching this for about 2 years like, "Hey, online fraud is going to go up. Here's what happened in other markets. Here's why it's going to happen. Here's what it's going to look like.” And it really did go way up during that time. And we saw fraud shift very drastically online. Um there was more fraud, there were a few more sophisticated attacks, but a lot of it was about volume. Um, and then shortly after that time or right around that time was when we started to see account takeovers which also dramatically changed the way that we fight fraud cuz then it was no longer just at the payment uh acceptance point but also at the, you know, the account login page. And that's where, you know, merchants and online companies were getting hit, you know, most. After they kind of got their cards under their card fraud under control and were able to identify that, then they'd see their accounts be hacked as what, you know, that's kind of the colloquial term. We say, account takeover. Um, so those are two examples of times when fraud changed in our world. This is another and that is when AI takes you know is starting to be developed. I so much has happened in the world of AI and their capabilities just in the last six months but I don't know what's going to happen in the next year the next time I'm presenting at MFA like I I just I can't even imagine. So I'm going to try something new and I'm going to share my screen and those of you on YouTube will be able to see my slides. Uh those of you not on YouTube, I will do my best to talk through them. That will be not so much of a challenge for me, but something just for you to be aware of.
Karisse Hendrick
04:46
If you're listening, you know, on Spotify or Apple Podcasts or whatever podcast app you use, feel free to switch over to YouTube. You can just search Fraudology. Uh you should find me right away. Um it might say the Fraudology podcast or Fraudology with Karisse Hendrick. I'm not exactly sure how my editor set it up. Um, but you'll find me. So, uh, that this is your opportunity to change that if you want to. Otherwise, we're just going to cruise right along. And I even have a video clip from the dark web that I'm going to try to show as well. So, might be a little clunky, so bear with me. Uh, but I did test this out earlier and it worked. So, I'm thinking it's going to work now. All right. Here I go. Sharing my screen. All right. So, here I go. Uh, really the takeaway from this session was that fraud has been industrialized. The title of the session was the fraud attack landscape. But, uh, really, you know, fraud has been industrialized. AI has lowered the cost, skill, and time required to attack e-commerce systems at scale. So fraud activity is accelerating. We have early indicators that point to faster growth across several e-commerce attack types such as uh I think it was from a signified state of fraud report in 2026 that fraud pressure has gone up 33% year-over-year from uh I think it was from 24 to 25 but I could be wrong. Um account takeover is up 78%. Card testing is up 175%. And by online pickup and store fraud is up 65%. Uh all of those things are because of AI. They're able to do things faster and cheaper than ever. And I think anyone that's been in fraud for a little while knows that it's impossible to have the goal of eradicating fraud on your website 100%. But the goal of you know most online merchants as well as banks and you know other online uh entities you know lenders crypto anything like that. Uh their goal in the fraud department isn't to make fraud go away 100%. They know that's not possible but it's to make it harder, make it slower, and make it more expensive to commit fraud than at other places. It's kind of like, you know, that saying, your job isn't to outrun the bear, it's to outrun the, you know, slowest person, uh, running from the bear. Um, and that's really what it is about fraud, right? You want to be more expensive and harder to penetrate and commit fraud on your site than other other websites, so that they don't attack you, they unfortunately attack them. Um, that's why I'm such a big believer in collaboration and why I worked so hard on the merchant fraud alliance conference, why I do Fraudology, why I host merchant calls for free. Um, it's all because, you know, if we work together, then we'll all be able to outrun that bear, right? Um, that's the goal. So AI is making it. So while that's our goal is to be, you know, more expensive, slower, harder to, you know, commit fraud on your website, AI just cuts all that down. It makes it faster. It makes it cheaper. It makes it easier to commit. No longer is it just one person at a screen or writing a script where you know every 10 20 seconds maybe you know a couple of transactions go through here and there but scripts are different than AI because AI is actually adapting to your website. It's determining, oh, it won't let me do this. Will it Will it let me do that? And so, it's constantly adapting to the things that you put in place to stop fraud. That's different than a script that's just going to continually do the same thing over and over again. And you can catch those, right? It's much more difficult to catch AI.
Karisse Hendrick
09:23
One of the reasons why account takeover is higher is because humans, your consumers are much easier to fish. And when they fish, your consumers, they get their login and their account details. So they can then log into your website. Um, fishing attacks using AI are so sophisticated now. They can even be customized. I don't know if you saw the episode or listened to the episode uh with Cy Khormaee who used to work for Google Recapture, but he was talking about just how you know easy it is to tailor a phishing message to a person. All you have to do is ask Chat GPT to do a little research for you. You know, you can find out where their kids go to school. You can find out, you know, where they work, what conference they're going to go to next. You can find out all kinds of things. And so you can tailor your phishing message to them. Lots of times you don't have to. A lot of the phishing messages these days are, you know, coming from what they say are large companies that everyone does business with or that most everyone does business with. So it doesn't look too out of the ordinary. If you were to get an email from some, you know, ABC, you know, cybershop.com or whatever, something really small, and said, "Hey, you need to log into your account." You'd be like, "What? I don't have an account with them." But you get it from a shipping partner or from, you know, Amazon or Walmart or PayPal or, you know, any big brand. Chances are you have an account with them. And so, you know, if it says you need to log in for security reasons or whatever they're telling you, you're going to give up your username and password. And that's why ATO's are up 78%. Card testing, again, same kind of thing. Instead of running a script, you can just feed hundreds if not thousands of rows of data of full card numbers into um ChatGPT or Claude or anything else.
Karisse Hendrick
11:25
I mean, I certainly wouldn't recommend doing that as a merchant, as a good guy. But the fraudsters don't seem to care. Um, and they can have them, you know, run transactions for a low dollar on a site very quickly and they'll always find ways to circumvent your fraud tools. I thought this was an interesting stat from Interpol. Um they estimate that the return from AI powered schemes versus traditional attacks to cyber criminals is four and a half times. So, if the average cyber criminal, you know, was able to obtain $1,000 from your website or your bank or whatever it is, in the average attack, in the traditional attack, they're now able to get, you know, $4500 worth of product or money or whatever else because AI is constantly learning. It's constantly adapting. It's going to get through. The next slide uh talks about how fraud operates as a connected system. When you're using uh AI, it's going to act as a connected system with everything else. So it's it's no longer just one attack. It's all the way across all attack surfaces. So identity abuse, payment fraud, support manipulation, and disputes all reinforce one another across the channels. So AI compresses the timelines um the time from research to attack. It allows coordination. Agents link tasks that once needed separate people and tools. And then it removes friction. Automation removes the trial and error that slowed traditional campaigns. The same campaign campaign can move from identity to checkout to customer support and dispute claims all within one campaign. So they can, you know, spoof an identity. They can successfully check out. Then they can go to your customer support and claim that the item wasn't received or claim that there was an issue or they were double build or whatever they claim, right? And then they can go to a dispute and either file a dispute internally with your company or with their bank all in one attack. It used to be that you know each thing was specialized by someone different within a cyber crime ring. You know, you had the hacker and the coder. You had the person that was good at identity, the person that was good at account compromise, the person or the group that was good at card testing, and then the group that was good at cashing out, as they say, or carding, you know, the checkout piece. Then they had another person or group that was good at manipulating and social engineering customer service. And then they'd have another group that would maybe file a chargeback or two if it was a fake card or something like that. Um, it'd all be different groups within maybe maybe it would be the s a single crime ring or it could be different crime rings and they're selling their services to each other through fraud as a service. But now using AI, it's just seamless. It the same agent, the same campaign can manipulate every single part of your business. Um, and that's really something that I, you know, wanted to highlight in this in this session. So, evidence is now cheap to fake. Um, if you're watching this on uh YouTube, just ignore that black box cuz I'm going to show you a video. I wasn't able to embed it because of the way that the um presentation was done through the AV system at MFA. A little behind the scenes there. Um, but I just put like a, you know, a a box there to in input that or insert that video, but I ended up not needing it.
Karisse Hendrick
15:34
Um, so evidence is now cheap to fake in a lot of different ways, right? They can fake receipts or order details or damaged photos. They can fake dispute narratives or voice calls or identity documents. Can all be produced on demand. So, no longer is a receipt from your company good enough for a return, right? If for an in person transaction, uh, no longer is are order details available, right? Like, they're all fakeable. They can fake damage photos. There was one of the largest uh food delivery companies uh in the US uh and they also have a pretty strong European channel as well. Um at MFA who ran a deep dive. I explained what deep dives were on the last episode, but basically it's a merchantonly uh session for 20 people for an hour where the facilitator shares a little bit of what's going on in their world um on a specific subject and then merchants get to talk and share ideas and share frustrations and all of that with each other for the rest of the time. And in his, you know, in his world, one of the biggest things to impact refund fraud is is damage photos, right? People can ask chat GPT to, you know, like they'll take a picture of something and then they'll say, um, add mold to this picture and it'll look really lifelike or, you know, add, um, you know, open the package or something like that. Or like, you know, if they if it was groceries, right? Crack all the eggs in this 12-pack or crack five of these eggs in this 12-pack and make it look like they're cracked. Oh, they can do that very easily, right? So, um that has really impacted the food delivery industry. It's also impacting uh refund fraud for retailers where you know you buy a piece of clothing, you can ask ChatGPT or Claude or you know, Anthropic, whoever you're using um to insert a hole in the in the product, right? So that you can contact customer service. You can have your AI agent contact customer service say there was a hole in the item that you shipped to me.
Karisse Hendrick
18:05
I'll send you a picture. It'll look lifelike. Um, you know, there was a lot of talk at uh in the session right after mine at MFA about faking identity docs. So, I kind of left that alone. I mean, I mentioned it, but I didn't dive into what deep fakes look like and how sophisticated they are. Uh, one of the slides that the next speaker shared was uh about just showing the difference between a deep fake video a year and a half or two years ago to last year to this year. And like two and a half or you know, one and a half to two years ago, the video was glitchy. You could usually tell, you know, they only had four fingers or like, you know, just weird things. The lighting was off, like you could tell things, right? And then last year it was a little little glitchy still, you know, not as glitchy as before, a little more seamless. Um, and the lighting was, you know, a little detectable, but not really. It was difficult to detect. Now, it's almost impossible to detect deep fakes. It's almost impossible to detect selfie videos or uh if a driver's license or passport has been forged. It's just not possible. Um so, three in 10 retail fraud attempts are estimated to be AI generated. That's probably going to be six or seven and 10 next year. Um, I'm going to go through the rest of this slide and then I'm going to do my best to play this video. Um there was a 20 there was a campaign done by AI that was done this year in 2026 where a 100 plus retail sites were compromised all at once with basically online skimmers card skimmers and six over 600,000 card records were stolen and they pocketed about $25 per attempt.
Karisse Hendrick
20:14
So do the math right? That's going to be over around $250,000 that they were able to get very quickly by implementing an AI attack across over a hundred merchants without the merchants being aware. Okay, we're going to play this video and basically it's Matt Vega created or created this video and it shows him going to this site on the dark web. Looks just like a regular site. It has rules of how to use it. Um, and then you can search any cards you want by the country, any country that you want. Uh, then you can search by um if it's credit or debit, what branded card it is, if you have a specific bin, but you can also pick the state that it's in. Uh, in this case, Matt picked Washington. Um, he put in the zip code. So, it can be very specific, right, to where he wants the item shipped. He can buy a card that has a high limit, doesn't have verified by Visa. Like, it has all of these details that you can see on here. Um, and then he went into the Russian market on the dark web and uh entered his pin and then he was able to um do even more things as far as um creating a FOSS, you know, that's what fraudsters call it when they have a full portfolio. Um, you can pause here and and read all of that text if you want. Um but it a gave them access to Imperial market where you could buy more clone cards, you could buy prepaid cards, you could buy Western Union, Skril, Crypto, etc. So, uh you can even buy Revolute transfers, uh which were um quite popular at the time. Um, and it would only be I think I saw a transfer for let's see $2,000. Oh, I wasn't able to see how much it was, but I mean it's pretty if it transfers for 2,000, it's probably going to be like $100 to purchase.
Karisse Hendrick
22:35
Um, so you know, this just shows that it's cheaper and easier than ever to commit fraud and to fake information um, more than ever. I mean, you can pick here. He's picking the bin. He's picking the, you know, the quantity of cards, how he's going to pay, etc. So, AI scales familiar fraud methods, right? So, we're not talking about new fraud methods. We're not talking about, oh, there's these new fraud methods. No, they just make it they make the current fraud methods that we have harder to identify as fraud. They make them blend in more with legitimate cards and legitimate transactions, legitimate account logins, legitimate account creations. They make it look better, right? So the um as I just said, the techniques existed before generative AI. AI just makes them cheaper to produce, test, and adapt. So, you know, there's under identity, there's stolen stolen data and deep fakes that create convincing identity replicas. Um, for ad traffic for fraud, for ad fraud, um, automated ad campaigns can send consumers to fraudulent storefronts. So, um, I've mentioned this before, Matt Vega has mentioned this before on my podcast as well, but you can make a very, very lifelike website for un under 10 within under 10 minutes for like under $50 online using AI. You can literally ask one of the chats to create a replica website for you of X do or not X.com because that's Twitter now, but company X.com. Um, you can replicate their storefront but have their credit card transactions not go through. Instead, whenever anyone enters their credit card data, you just harvest that and then they don't ever get their product. But you get their credit card numbers. Um, or you can try to get a merchant processing account uh fraudulently of course and probably using AI because it's easier now and link that to your fake website so that whenever someone places an order, you are processing their card for that amount, but they'll never get the merchandise back. So then with this ad traffic, you can then create hundreds or thousands of ads on social media pointing to your fake website. So when you do that, they don't usually see the URL. So you could have the URL be spelled wrong with just one letter or something like that. And if they're clicking on an ad from social media about the website that they think it is, they're not going to notice that you're that it's not the real website. So that has been greatly impacting online companies, including banks. Um phishing, as I mentioned, you can target research and polish messages um making them uh a lot more convincing.
Karisse Hendrick
26:04
You can Oh, this is what I just talked about. Copycat shops can be built quickly to collect payment data or sell counterfeit goods. Just like I said, card testing, the authorization results help sort cards, bins, and address matches. So, when a card tester uses AI, it doesn't just tell them if the transaction has been authorized or not. It'll sort them based on BIN. It'll sort them based on what the authorization, you know, response was. It can sort them by the address so that you can, you know, add add cards that are close to your mules address or, you know, purchase digital goods that don't need a shipping address. And then the biggest thing is that it adapts to everything. So agents map friction and controls and then they adjust their approach. So if they hit a brick wall, if they hit, you know, a fraud tool or fraud screening, they then make note of that and try to figure out how to get around that. And you're not seeing as many trial and errors in your data, right? Especially if they're using card testing on another site, you're not seeing, you know, 2700 declines and 3,000 authorizations or or 300 authorizations, right? That would tip you off the other card testing problem. But now they're card testing somewhere else so seamlessly that when they go to use the card number at your website, it's really difficult if not almost impossible to solve for or at least it's not impossible but we need to be thinking about it in a different way and that's that's really what a lot of this is about. Um so the losses surface across the life cycle right I basically said that but to capitalize on it you know a campaign can run across controls before a team sees the pattern right so um the discovery is the fake stores and ads and then they use stolen or synthetic ID to commit card testing and checkout. Um, the account can be used for takeover or misuse. Um, often times they're taking over an account to drain all the loyalty points or the stored value dollars that you have on the account. If you're a seller for a marketplace and they get a handle a hold of your account details, they're going to go in and instantly change the payee or who the money is deposited to so that they can then get that money instead of you. Um they will, you know, manipulate support with fraudulent refund claims and then they'll go to this dispute process if they need to and issue a chargeback uh with the bank and they can provide flawless card holder documentation because it's all fake.
Karisse Hendrick
29:13
Um there's a reason why eventbased controls miss it, right? And this is what I was just hinting towards a minute ago. Approvals, login, and authentic identities can all look valid in isolation. The loss emerges later after all after many successful transactions. So often times you won't realize that you're under attack until after you've already been attacked. Whereas with previous fraud attacks without AI, you'd have a pretty good idea you were being attacked before you get a chargeback or before you get a card holder calling your customer service to tell you that their card was used on your site. Talked a little bit about refund abuse already, but I'll talk about it again. Um, 60% of Well, wait. So, AI can generate uh damaged photos and polished claims. Uh, deep fake refund bots can probe call centers around the clock. So that's another thing. They have deep bake refund bots that will continually reach out to customer service to get a refund. And if it's a 24-hour call center, which a lot of them in the US aren't anymore, but if it is, they're going to catch someone at 2:00 in the morning that's absolutely exhausted. Or they're going to look on LinkedIn. You know, AI can look on LinkedIn for everyone that works that has the title customer service rep at X company or XBank and then, you know, Crofts reference that with their Facebook accounts to see who most recently had a baby. So, if you're calling them at 2:00 in the morning and they're on the call center, probably they're probably exhausted. So, that's a great time to call customer service and fake, you know, a refund claim. Um, Europe and the UK said that, um, account takeover attempts have gone up 67%. And items not received have gone up 49%.
Karisse Hendrick
31:19
And items not as described have gone up 35%. That's just in the last like year. That that's looking behind. Looking forward, it's going to be going up so much more. Um, this is what So, what can you do about it? Right? I'm never going to present a problem without some kind of a solution. And the teams that are doing a really good job at catching AI attacks are the ones that are connecting identity, intent, and context from discovery through post purchase. So, they're able to read patterns across sessions and transactions from the moment a user enters their website all the way through checkout and then maybe even after checkout. Um, they're looking at early signals, right? They'll they're able to spot misuse or disputes or account compromise before the losses get too high. So, back to what I was saying there. So, what I've noticed is a lot of the crypto companies did this about two to three years ago where they now look at the user journey. They're not just looking at events, right? Account creation, account login, payment p or purchase or transfer or withdrawal. They're not just looking at those things. That's kind of like taking a snapshot at the time like a still picture of someone breaking into your house, right? You can know that someone broke into your house, but you don't know what they took. They don't you don't know where they went in your house. You may not even know how they got into your house if it's just a still picture. But if you have security cameras outside and in, you can then see how they got into your house in the first place. Did they break a window? Did they find an open window? Did they find an open door? Did they find Did they clone your garage door opener? Whatever it is, right? If you have 247 security cameras inside and out, then you're better off being able to um not only identify how they broke in so no one can break in again, but also what they took, what where they went first, what their actions were, how they were able to get past your safe or whatever else you have. Um that's the difference between using point solutions that are measuring things at specific times and and you know taking that snapshot at the time of purchase, the time of login, the time of account create, whatever it is. Um the time of transfers or cash outs or things like that. Um newer services that allow insight into the customer journey um that provide session data rather than just transactional data um and are looking at the sessions all the way through are then able to pick up on those patterns that fraudsters don't realize that they leave or as Soups Ranjan said in a session that I moderated with him every fraudster has a tell. And when you're able to monitor all the way through from the beginning of when they enter your website all the way through to the end when they check out or when they withdraw money or whatever they their action is, they will leave a tell. They will have their phone upside down. They will, you know, and if you're able to look at telemetry through your device, you know, your digital um fingerprint or your device ID service, then you're able to know, hm, it's impossible to shop on my website when their phone is face down. Or, you know, you can know how they hold the phone or even one vendor was talking about apartment number location number uh precise like location. Um, I think they do that using telemetry of the phone and how many feet or meters it is in the air. Um, but and so they'd be able to tell you what floor uh someone is on when they're making a purchase or making a withdrawal. Um, and if that matches the apartment number. So, if they, you know, live at 1206, apartment 1206, you could know that they were on the 12th floor when they made the purchase. Will that make sense? That type of thing. There's all different data signals that different providers will provide. Sorry, that was a bit redundant. Uh different providers will give you, but the important thing is is that you are ingesting all of those and that you're looking at those patterns.
Karisse Hendrick
36:15
You're looking for those tells. Um then one of the things I wanted to leave I wanted to give kind of a note to the solution providers that are in the room the sponsors. This stat came from the Merchant Risk Council's uh fraud survey and it says that 80% of merchants name technological infrastructure as their biggest fraud challenge. So their biggest fraud challenge is an account takeover. It's not card testing. It's not, you know, getting budget for operations or anything like that. It's getting the right technological infrastructure. That's what their biggest fraud problem is. And that's a hurdle you have to help, you know, the decision maker or the influencer, you know, get across the finish line with their leadership because, you know, you're not going to see it all the time. You're not going to know it exists. Um but yeah, so that's, you know, that's the biggest hurdle and I think that's really good. Um, speaking of signals, you know, looking at for those early signals, looking for those common signals across multiple uh transactions or actions done by fraudsters. Um, the last thing that it says on this page is that a missed signal can become a fraud loss, a chargeback, or a false positive. And that is 100% true. You know, just one missed signal. So, you need to ensure that you're ingesting the right signals. You need to ensure that you're using the right products because not all products are looking at session level data yet. They're not looking at device ID and behavior biometrics in combination. Uh those are the things that seem to work for the top leaders in fraud. The groups that you know continuously have very impressive authorization and decline rates, very impressive chargeback rates that are low. Um very little false positives. So, you know, that's how you get there. I thought that that would be helpful for me to walk through and explain where we're at when it comes to AI and technology and why it's changing so much. So, if you're having to present to your leadership or you whether it's executive leadership or your, you know, direct leadership about why fraud has increased so drastically. Uh, I know of several merchants that since April or May have had to be had to explain to their leadership why their account takeovers just went skyhigh all of a sudden or why their card testing just went through the roof and their credit card processor is now contacting them and talking about the enumeration program from Visa or, you know, why they have such high refund rates or chargeback rates. Due to item not described or not delivered. Um all of these things, you know, are when you're in the room presenting those things, remember this presentation. Remember the fact that AI is is the use of AI by fraudsters is going up very rapidly. It's increasing very rapidly. Therefore, fraud instances are increasing very rapidly. I hope that that for you that doesn't translate into fraud losses going up significantly, but it could. And the best solution is more tooling, you know, more visibility into the entire customer life cycle, the entire customer journey. And by the way, if you have user journey data, you know what department really loves that journey data? Marketing. Marketing loves to know where they're, you know, what their customers are clicking on, what they're doing. So you can provide some you know really good intel to marketing or your C suite from those products because you know mapping it all out is one thing but being able to translate it to leadership to hopefully get increased budget to increase the capabilities of your tools. That's where things are going to change.
Karisse Hendrick
40:59
That's where we're going to see. Yeah, attempts are high, but the losses aren't because we were ready because we were covering more attack surfaces than we ever have before. So, that is my presentation. Um, just almost exactly 45 minutes. Uh, on stage it was about 30, but I had a countdown clock. I didn't have a countdown clock this time. Um, but maybe I need that to work that into my podcast. I don't know. But uh let me know what you think. Let me know if you disagree with me. Let me know if you agree and have an example. Um this is what I'm seeing from a 10,000 foot view of fraud. Uh when I talk to lots of companies and that is that fraud is getting cheaper and easier for fraudsters to commit. Therefore, fraud attacks are going up. You know, a 33% increase in fraud pressure is not insignificant. Um, and feel free to use any of these stats that I compiled. I compiled them from various studies uh that have been available in the last six months. Um, feel free to use them to present to your leadership and say, "Hey, this is what's happening at the industry level. Fraud is going up." So the fact that our fraud is going up shows that we're on par with the industry, but it also shows us that we need to change our tooling because our current tools AI is able to figure out where the vulnerabilities are and exploit the hell out of them. So that is my public service announcement of the day. Let me know what you thought of the presentation. If you like seeing me have slides, if you like the visual, uh, other than just watching me talk and stare into my laptop, um, let me know what you think about the presentation and the data. Um, if you disagree with it, I would be curious to know why. Uh, but I can tell you in talking with all of the merchants that I do, and I mean, it was literally hundreds just last week at the conference, everyone's feeling the pressure. And on the flip side, when I attended the AI boot camp that I talked about on last week's episode, the general consensus was that because of AI, leadership is expecting you to do more with less.
Karisse Hendrick
43:22
They're expecting higher sales and bigger growth and all of that, but they're expecting you to figure it out and figure out how to do more with less. So, you know, I do I am a big proponent on getting the right technology in, but there are people that are being really scrappy right now. There's one merchant that was having an issue with phishing messages in their chat. So, they allow their buyers and sellers to chat with each other about the product and that type of thing. And they were finding these fraudsters taking over legitimate accounts and then going into chats with um as a prospective buyer with sellers to say, "Hey, uh I'm really interested in what you're selling, but I need to have you, you know, click this link to contact me or whatever it is." Somehow they get them to click the link. Once they click the link, they provide their information and then they can take over their account and steal the stored value that's in it. What this one of the merchants, they've been doing a lot of different creative things using AI trying to identify these, but one of their analysts who is one of the smartest analysts I've I've met in a very long time and they may be looking for a new job. So if you have something remote in the US, let me know. But very just so smart and thinks so out of the box. They've been able to use chat GPT to say so they've loaded in all of these messages like the you know the way that the message flows you know and all that into a chat GPT or Claude or whoever they use along with good message you know back and forth between buyers and sellers and say hey which ones of these look like they were done by AI or and then they say okay of the ones that were done by AI. How can I identify these in the future? And one of the things they found was that every uh chatbot starts with they make the first move right in the chat and they say like, "Hi, how are you?" And then you know the person they ask an open-ended question, right? And then or maybe it's about the product or whatever. And then the, you know, good seller, the legitimate, the real person, the seller, it answers back. And then they ask another open-ended question. And then the seller answers back and then they prompt for the link. So there's like a very specific they've also looked at the words they use. They've looked at the syntax that they use. And interestingly enough, AI can identify AI. It could identify when AI is is, you know, writing out sentences. So, that's just one example of things that merchants are doing to be scrappy to try to stop this stuff. And it's working. There's lots of really interesting things that merchants are doing with AI to try to stop this generative AI or LLMs or, you know, if they have internal AI with that's paired up with their data internally, you know, like a private account that's paired up with their and a private model that is paired up with their data um that they're able to play around with. They can do all kinds of analysis that they may not have thought about before. So, there are some hidden silver linings, but if you're expected to do more with less, and fraud is going to be just increasing at an all-time high over the next several quarters, you're in for a busy time. Um, and that is why we did the AI boot camp because we know people are being expected to use AI in specific ways um or in in specific ways where they don't tell you how to use it. They just say, "Well, you need to reduce your human time by 25% or whatever." So, we're living in very interesting times, but we can fight AI with AI.
Karisse Hendrick
47:39
There are um providers out there that have AI agents that will screen, you know, manual review for you or that will look at all of your rules and see which ones fired and propose new rules to write or new things for the model to ingest. So, we we have a few tricks up our sleeve, too. But there's more of them than there are of us. So, it's really good for us to share information and that's ultimately why I wanted to share this presentation and why I do this podcast. So, on that note, I am going to go for the day. I hope I gave you a few things to think about, a few nuggets of information to ponder over. And don't forget to reach out to me if there was anything significant in this uh I was going to say session, but this episode I at the conference I was calling them episodes and now on the podcast I'm calling them sessions. I just my brain gets mixed up because I really enjoy creating content. But you know, if it gave you fuel for thought, if you have more questions, I'd love to hear from you. It might turn into a second, you know, a follow-up podcast episode or I might just write you back. Uh, LinkedIn is preferred over my email. Um, I don't get as much spam in my LinkedIn as I do in my email. Um, and especially with the political season and the midterms happening in the US, I don't know who I donated to and when, but man, am I getting a lot of those emails now. And I stupidly used my work email address. So things are getting lost, but that's another story. Uh, I'm going to let you go for the rest of the day. I hope I gave you some things to think about, and I will talk to you more next week.
View full transcript
