Outtake Raises $40 Million to Track Impersonation Scams

U.S. AI security startup Outtake raised $40 million in January, bringing its total funding to about $60 million. The company uses AI agents to find fake websites, accounts and phishing content impersonating businesses, trace related attack networks from suspicious sites, and automate requests to platforms to remove them. Palantir alum Alex Dillon founded Outtake in 2023. James Oh, a software engineer and head of experimental product development, said the company’s core strategy is to use AI to stop threats before users fall for impersonation attacks.

Oh said AI is accelerating both vulnerability discovery and offensive and defensive work. In April, Anthropic disclosed the security vulnerability-finding capabilities of its Claude Mythos Preview model and began reviews with major software companies. Software vulnerabilities can be addressed with patches, Oh said, but technical patches alone cannot stop attacks that persuade users to trust fake sites or impersonation messages and hand over information. Social engineering attacks exploit human trust and judgment and include business impersonation texts and phishing. Generative AI has made it easier and faster for attackers to create fake sites and content.

Outtake focuses on finding threats early and following through with removal requests, rather than only helping users identify fake content. It gathers information from public websites and domains, app stores and social media. Its AI agents analyze clues in text, images, video and audio to classify possible business impersonation or financial fraud. Multiple agents divide the work of gathering information, classifying risk, conducting deeper investigations and responding; the arrangement connecting them is called agent orchestration. Outtake’s Recon Agent starts from a suspicious domain or account, tracks related websites, accounts, IP addresses and attack infrastructure, and writes investigation reports. It also checks whether the same scam recurs across platforms.

The company’s main customers are technology, financial services and consumer goods firms, as well as individuals such as prominent influencers and entertainers. During Cash App’s first three months using Outtake, the startup reviewed more than 3.5 million video-focused social media posts and found more than 2,400 pieces of scam content. While older tools mainly examined text and images, Outtake also analyzed speech and audio in videos. OpenAI is also a customer. In July last year, Outtake described a case using OpenAI models to detect and classify threats and automate responses. In July this year, Anthropic disclosed how it developed the Recon Agent using Claude. AI does not make every decision or directly remove sites. Customers set the brands and accounts to protect and the criteria for decisions; security or legal staff can review difficult cases. The platform decides whether to remove a site or account after receiving a request.

Outtake is also studying attacks that try to deceive AI agents themselves. The concern is that agents reading the web, collecting information or carrying out transactions may mistake instructions in outside posts for work commands. In February this year, Oh investigated Moltbook, a social network for AI agents. OuttakeSheriff, a security agent, analyzed 106,097 posts and selected 134 suspicious ones; a separate verification agent classified about 20 of them as actual threats. The first agent broadly flags suspicious cases, then the second reviews their content and context, reducing the chance that security discussions are mistaken for attacks.

The investigation found a case where instructions to follow an account were hidden in HTML comments on a webpage and invisible to people. It also found posts encouraging token transfers or email sending. Prompt injection is a technique that hides malicious commands in external content to alter AI behavior. Investigators found no cases of information or money actually being stolen, but the findings showed that agents may wrongly treat internet content as legitimate instructions. Oh said model improvements alone cannot eliminate the risks. Suspicious material from external webpages should be filtered so agents receive only necessary information, and their permissions should be limited. Outtake is also developing a way to isolate the execution environment so agents investigating external sites cannot access sensitive internal systems.

Oh said distinguishing genuine services from fake ones will become important as AI agents do more work on the internet. He viewed the Korean security market’s growth potential positively and predicted that as AI lowers the cost of developing and using software, demand for security to protect it and verify authenticity will also grow. He added that the scale and speed of attacks have already changed, so the security industry must respond to AI-powered attacks and defense companies must learn how to use AI.

U.S. AI security startup Outtake raised $40 million in January, bringing its total funding to about $60 million. The company uses AI agents to find fake websites, accounts and phishing content impersonating businesses, trace related attack networks from suspicious sites, and automate requests to platforms to remove them. Palantir alum Alex Dillon founded Outtake in 2023. James Oh, a software engineer and head of experimental product development, said the company’s core strategy is to use AI to stop threats before users fall for impersonation attacks.

Oh said AI is accelerating both vulnerability discovery and offensive and defensive work. In April, Anthropic disclosed the security vulnerability-finding capabilities of its Claude Mythos Preview model and began reviews with major software companies. Software vulnerabilities can be addressed with patches, Oh said, but technical patches alone cannot stop attacks that persuade users to trust fake sites or impersonation messages and hand over information. Social engineering attacks exploit human trust and judgment and include business impersonation texts and phishing. Generative AI has made it easier and faster for attackers to create fake sites and content.

Outtake focuses on finding threats early and following through with removal requests, rather than only helping users identify fake content. It gathers information from public websites and domains, app stores and social media. Its AI agents analyze clues in text, images, video and audio to classify possible business impersonation or financial fraud. Multiple agents divide the work of gathering information, classifying risk, conducting deeper investigations and responding; the arrangement connecting them is called agent orchestration. Outtake’s Recon Agent starts from a suspicious domain or account, tracks related websites, accounts, IP addresses and attack infrastructure, and writes investigation reports. It also checks whether the same scam recurs across platforms.

The company’s main customers are technology, financial services and consumer goods firms, as well as individuals such as prominent influencers and entertainers. During Cash App’s first three months using Outtake, the startup reviewed more than 3.5 million video-focused social media posts and found more than 2,400 pieces of scam content. While older tools mainly examined text and images, Outtake also analyzed speech and audio in videos. OpenAI is also a customer. In July last year, Outtake described a case using OpenAI models to detect and classify threats and automate responses. In July this year, Anthropic disclosed how it developed the Recon Agent using Claude. AI does not make every decision or directly remove sites. Customers set the brands and accounts to protect and the criteria for decisions; security or legal staff can review difficult cases. The platform decides whether to remove a site or account after receiving a request.

Outtake is also studying attacks that try to deceive AI agents themselves. The concern is that agents reading the web, collecting information or carrying out transactions may mistake instructions in outside posts for work commands. In February this year, Oh investigated Moltbook, a social network for AI agents. OuttakeSheriff, a security agent, analyzed 106,097 posts and selected 134 suspicious ones; a separate verification agent classified about 20 of them as actual threats. The first agent broadly flags suspicious cases, then the second reviews their content and context, reducing the chance that security discussions are mistaken for attacks.

The investigation found a case where instructions to follow an account were hidden in HTML comments on a webpage and invisible to people. It also found posts encouraging token transfers or email sending. Prompt injection is a technique that hides malicious commands in external content to alter AI behavior. Investigators found no cases of information or money actually being stolen, but the findings showed that agents may wrongly treat internet content as legitimate instructions. Oh said model improvements alone cannot eliminate the risks. Suspicious material from external webpages should be filtered so agents receive only necessary information, and their permissions should be limited. Outtake is also developing a way to isolate the execution environment so agents investigating external sites cannot access sensitive internal systems.

Oh said distinguishing genuine services from fake ones will become important as AI agents do more work on the internet. He viewed the Korean security market’s growth potential positively and predicted that as AI lowers the cost of developing and using software, demand for security to protect it and verify authenticity will also grow. He added that the scale and speed of attacks have already changed, so the security industry must respond to AI-powered attacks and defense companies must learn how to use AI.