By Rohan Gupta, Co-founder and CEO, Diopter AI
For most of the last decade, social engineering was a craft. An attacker researched one target, built a pretext, and worked the phone one person at a time. AI is removing that bottleneck, and I expect social engineering to be running at machine scale in 2027. Security teams will respond the way they did when malware outgrew antivirus, by putting live detection and response where the attack happens. For social engineering, that place is the conversation.
1. Social engineering will run at machine scale
In less than a year, the shift went from a theoretical risk to something showing up on incident reports. In November 2025, Anthropic disclosed an espionage campaign in which AI performed 80 to 90 percent of the work, with humans only stepping in at maybe four to six decision points. In August 2026, the UK AI Security Institute reported that an AI agent that was never told to deceive anyone created fake identities and used them to pressure a real open-source maintainer into approving malicious code during a routine cyber evaluation. The maintainer refused and AISI stressed that its test conditions were intentionally loose, but the agent still decided to try social engineering on its own.
Social engineering has always been limited by the time it takes an attacker to conduct research, design pretext, and make individual calls. Agents cut that time down significantly, so I expect 2027 to bring campaigns that target an entire workforce at once, with each attempt tailored to the person receiving it. We’re already seeing it happen. One hiring team told us it received 1,500+ applications in five days for a single analyst role and suspected most of them were fake.
2. Deepfake detection will become the standard, and the ask will become the signal
Deepfake detection will matter in 2027 and every serious security stack will need it, but it will also stay a cat-and-mouse game. Detectors are probabilistic, so they tell you how likely it is that a voice or face is synthetic based on the generators they were trained on. Every new generator makes that training less useful, and attackers make the signal worse on purpose by dropping video resolution or blaming a bad connection.
More importantly, many of the most damaging attacks don’t use a deepfake at all. The FBI and CISA describe Scattered Spider posing on both sides of the help desk over the phone, talking employees into running remote access tools while impersonating help desk staff, and manipulating help desks into resetting passwords and MFA tokens. Once they were inside, the group hunted for VMware vCenter, and they’ve been observed encrypting ESXi servers. The road to the hypervisor ran through a help desk call, and the advisory notes the group also targets contracted help desks, a risk for every organization that outsources support to an MSP.
Each attack shows a consistent pattern of claimed authority, a sudden deadline, a move away from the usual channel, a request for secrecy, and finally the ask. That ask might sound like “reset this MFA,” “install this tool,” or “change this bank detail,” and that’s where detection can get close to deterministic, because a policy is a rule. A password reset that doesn’t follow the right process is an issue whether the caller is a malicious agent or a skilled human.
So the question buyers ask will shift from “Is this person real?” to “Should this request be happening at all?” The FBI’s Internet Crime Complaint Center recorded $3.05 billion in business email compromise losses in 2025, but businesses tied just over $30 million of those losses to AI. I suspect the gap exists because people only report the AI they actually notice, and a defense that depends on seeing it will miss much of the rest.
3. Security teams will treat conversations like endpoints
Endpoint security used to mean signature antivirus, until EDR gave security teams live telemetry and a way to respond mid-attack. The human layer is still in the antivirus era. A security team can see nearly every process on a managed laptop but usually nothing on a call where the help desk resets an executive’s MFA.
Training fills that gap today, but buyers are candid about its limits. One IT director told us his awareness program is “really just for phishing. It’s not really social engineering.” A security leader at an investment firm said his team explains policy to every new hire, “but it would be nice to have a technical enforcement behind that policy.”
That is why buyers keep reaching for the phrase “EDR for voice and video.” In 2027, that category will evolve into live detection of synthetic voice and video, analysis of how a conversation is escalating, checks against the company’s own policies, and alerts routed to the security team with evidence attached. It is what we are building at Diopter, and we aren’t the only ones who see the need.
Privacy concerns will determine which of these tools actually get deployed. Recording and data retention are the top objections we hear, so I expect on-premises deployment and zero data retention to become purchase requirements in financial services and other regulated industries.
Going into 2027, the question for every security team will be whether it can detect a sophisticated social engineering attack in real time before access or money changes hands.
##
ABOUT THE AUTHOR
Rohan Gupta is co-founder and CEO of Diopter AI, which is building an EDR for humans: real-time detection of social engineering and deepfakes on live voice and video calls, so security teams can respond as policy violations happen. Before Diopter, Rohan founded QuillBot, the generative AI writing platform, and led it as CEO as it grew to more than 35 million monthly users.
