Polygraf AI announced Meeting Guard on July 14, a fraud-detection product that sits inside a video call and watches it happen. As reported by Luis Millares at Channel Insider, the tool joins meetings on Zoom, Microsoft Teams, and Google Meet as a visible participant and screens the session in real time for four things: synthetic voices, responses generated by an AI rather than a person, identity impersonation, and the live exposure of personal data that somebody says out loud without thinking about it.
The product runs on what the company calls its AI Behavioral Control Plane, which bundles deepfake voice detection, flagging of AI-generated responses, live detection of personally identifiable information, audio redaction, and a compliance audit trail. Polygraf AI says it requires no integration work and operates only while a meeting is running. The infrastructure is SOC 2 Type II and ISO 27001 certified, the company states a strict no-training policy on customer data, and a standard cloud deployment is quoted at under fifteen minutes with hybrid and on-premises rollouts running one to two weeks. Meeting Guard also functions as a notetaker, producing summaries without sending the audio to an external service.
Yagub Rahimov, Polygraf AI's chief executive, framed the problem in blunt terms:
Every meeting is now a security event. AI has fundamentally broken the trust model organizations relied on for remote communication. The ability to verify identity through a face, voice, or conversation is no longer enough.
Yagub Rahimov, CEO, Polygraf AI, quoted in Channel Insider, July 14, 2026
The Conditions Driving This
Most organizations have already been hit. Polygraf AI cites Gartner research finding that 62% of organizations have experienced a deepfake attack, with 37% of those involving video calls and 43% involving audio calls, which means the video meeting stopped being a hypothetical attack surface some time ago and nobody moved the defenses.
The dollar figures are now large enough for a board to notice. The company estimates annual exposure from AI-enabled meeting fraud running from $2.5 million to more than $71.4 million per organization, covering executive impersonation, vendor impersonation, hiring fraud, and financial scams, and it is worth reading that as a vendor's estimate rather than an audited number.
The hiring pipeline is becoming a delivery mechanism. Gartner projects that one in four candidate profiles worldwide will be fake by 2028, which turns every remote interview into a potential infiltration attempt and turns the recruiting function into a security surface that no recruiting team was trained to defend.
The controls enterprises already bought do not see this attack. Multi-factor authentication, endpoint protection, and identity and access management all verify that a legitimate credential is being used from a legitimate device, and none of them examine whether the human face and voice on the other end of the call belong to the person the credential says they are.
Biometric verification has become a target rather than a defense. Daniel Elliot, chief executive of the cybersecurity firm Delta Bear, argues that biometric checks can now be manipulated by attackers with the resources to reproduce body language, facial features, and hand movements, which means a control built on the assumption that a face is hard to fake is resting on an assumption that stopped being true.
The payment authorization call is the highest-value target in the building. A finance team that receives a video call from an executive authorizing a transfer has, until recently, treated the presence of the executive's face and voice as sufficient verification, and the entire approval chain in most companies was designed around exactly that assumption.
Cyber insurance has not caught up. Elliot, who has reviewed policies in this area, says the coverage available to small and mid-sized businesses is generally not written for AI-enabled attacks, which leaves the organizations least able to absorb a loss holding a policy that may not respond when the loss arrives.
What AI Security Looked Like Before This
Enterprise security has spent two decades hardening the perimeter, the endpoint, and the identity layer, and it did a reasonable job of all three. The email gateway filters phishing. The endpoint agent watches for malicious code. Multi-factor authentication confirms that the person logging in holds a credential and a device. Identity and access management decides what that credential is permitted to reach. Each of those controls answers a question about a machine or a token.
None of them answers a question about a person. The video call was treated as a trusted channel because, for the entire history of the technology, seeing somebody's face and hearing their voice was considered proof that they were who they claimed to be. That assumption was never written down as a security control, which is precisely why nobody thought to review it. It simply sat underneath the approval workflows, the vendor relationships, and the hiring process as an unexamined premise.
Deepfake technology has been discussed for years, and for most of those years it lived in the category of things that were technically impressive and operationally irrelevant. Producing a convincing synthetic person took skill, time, and money, and the effort was not worth the return for an ordinary fraud. What changed is that the effort collapsed while the return stayed the same. Elliot, speaking on Channel Insider's Partner POV, put the capability plainly:
I could definitely impersonate you. Matter of fact, I'm pretty sure I can impersonate you better than you.
Daniel Elliot, CEO, Delta Bear, speaking on Channel Insider Partner POV
What It Looks Like Now
Meeting Guard represents a category of control that did not previously exist in most security stacks: verification applied to the human channel while the conversation is happening. The design choices are worth examining, because they are the parts most likely to be copied by whoever follows.
The tool is a visible participant. It appears in the attendee list, which is a deliberate decision with two consequences. It makes the control auditable, since everyone on the call knows the session is being screened. It also removes the covert-recording problem that would otherwise make the product unsellable to a legal department. The tradeoff is that an attacker also knows the screening is running, though it is difficult to see what an attacker would do with that knowledge other than leave.
The detection runs across four surfaces at once. Voice analysis looks for synthesis. Response analysis looks for text that reads as machine-generated rather than human. Impersonation analysis compares identity signals. And a data-loss layer flags personal information as it is spoken, with audio redaction available. That last one is the least discussed and possibly the most immediately useful, because the ordinary case is not an attacker at all. It is an employee reading a customer's account number aloud on a call that is being recorded and processed by a note-taking service nobody vetted.
The deployment model is the part that will decide whether this category grows. No integration work, operating only during the meeting, and a cloud install quoted at under fifteen minutes describes a product designed to be bought by a security team that has already run out of budget and patience. The on-premises option, at one to two weeks, exists for the regulated buyers who cannot let meeting audio leave the building, which is the same constraint that shapes every other purchase those buyers make.
The broader shift is that trust in a live human channel is being converted from an assumption into a control with evidence attached. An organization that runs a screening layer on its meetings can now say what was checked, what was flagged, and what the audit trail shows, which is a materially different position from asserting that the person on the call seemed fine.
Our Take
AI Security Take
The underlying claim here is correct, and it is worth separating from the product that carries it. A face and a voice are no longer evidence of identity, and any approval workflow that treats them as evidence is running on a premise that expired. Detection tooling in the meeting is a reasonable response, and Polygraf AI is early to ship something specific rather than a warning.
Three things belong on the evaluation list before a purchase order.
The exposure figures come from the vendor. A range running from $2.5 million to more than $71.4 million is wide enough to describe almost any organization, and a range that wide is a marketing artifact rather than a risk assessment. The Gartner findings on deepfake attack prevalence are the harder number in the announcement and the one worth building a business case on. Ask for the underlying research before repeating either.
Detection accuracy is the question nobody in this category wants asked. A tool that flags synthetic voices will produce false positives, and a false positive on a live call with a customer or a board member carries a real cost. Ask what the false-positive rate is, ask what happens operationally when the tool fires, and ask who is authorized to end a call on its say-so. A control that nobody is empowered to act on is theatre with a subscription fee.
And a detection layer does not repair a broken approval process. The organizations that lost money to executive impersonation lost it because a single video call was sufficient authority to move funds, and that was a process failure before any AI was involved. Callback verification on a separately held number, dual authorization for transfers above a threshold, and out-of-band confirmation for any change to vendor payment details are controls that cost nothing, work whether or not the attacker is synthetic, and should already be in place. Buy the detection if the risk profile warrants it. Fix the workflow regardless, because that is the part that fails first and the part no vendor can install for you.
