Search an enterprise security budget for “deepfake defense.” It is not there. Search the fraud budget. Not there, either. Search identity and access management. Still not in many organizations’ security budgets.
U.S. identity fraud and scam losses hit $47 billion in 2024, according to Javelin Strategy & Research, and most enterprises cannot point to a single line item designed to stop the attacks that generated them.
The Deloitte Center for Financial Services projects that generative AI-enabled fraud losses in the United States could reach $40 billion by 2027 for banking and financial services alone — a separate figure tracking a narrower slice of the same threat landscape. Scamming software sells on the dark web for as little as $20. Voice cloning requires seconds of sample audio. The cost of impersonating a CFO has collapsed. Finding those losses in an enterprise budget is another matter. The budget line does not exist.
The budget blind spot
The Javelin figure primarily tracks consumer impact. The enterprise signal is in how those losses execute. For example, account takeover fraud alone hit $15.6 billion in 2024, up 23% year over year, and the techniques driving those takeovers, deepfake impersonation, credential theft, and synthetic identity creation, are the same ones now targeting enterprise onboarding, vendor verification, and financial authorization workflows.
FinCEN issued a formal alert in November 2024, warning financial institutions of rising deepfake fraud in suspicious activity reports. The agency requested institutions tag deepfake-related SARs with the key term “FIN-2024-DEEPFAKEFRAUD” precisely because these incidents were being lost in broader fraud reporting categories. When a federal regulator creates a new fraud tag, the losses have already outgrown the existing taxonomy.
Deepfake fraud incidents in Q1 2025 surpassed the total for all of 2024, according to Resemble AI’s Q1 2025 Deepfake Incident Report, with documented losses exceeding $200 million globally in a single quarter. The spend to stop it sits with anti-fraud teams. The threat sits with identity infrastructure, and the gap between those two line items is where attackers operate — nobody owns it.
Kayne McGladrey, IEEE Senior Member, told VentureBeat that the invisibility is structural. “If you can’t say, ‘if we had a deepfake in our accounts payable process, here is the material loss,’ then you’re not going to get budget for technology to solve the problem, because nobody cares about the problem because it doesn’t cost money.”
That is not a failure of technology. It is an accounting failure.
Why identity stacks break against synthetic humans
The CrowdStrike 2026 Global Threat Report found that 82% of detections in 2025 were malware-free, continuing a steady climb from 79% in 2024 and 51% in 2020. No payload or exploits were needed. Attackers logged in with stolen credentials and blended with legitimate traffic. Breakout times have compressed to minutes, with lateral movement beginning before most security teams can triage the first alert. AI-enabled adversaries increased attacks by 89% year over year.
KnowBe4, a company whose entire business model is security awareness training, publicly disclosed in July 2024 that it had unknowingly hired a North Korean operative who attempted to load malware the moment the company workstation arrived. The operative had passed four video interviews, a background check, and successfully passed a confirmed photo match. Operatives are clearing every gate because the gates were designed to verify credentials, not humans. If a firm built around detecting social engineering cannot catch a synthetic identity in its own hiring pipeline, the gap is architectural, not educational.
By 2027, 50% of enterprises will invest in disinformation security products or services and TrustOps strategies, up from less than 5%, according to Gartner. That gap maps the distance between where enterprise defenses sit and where they need to be. The question is whether budgets move before the losses force them to.
Manual workflows against machine-speed attacks
The speed mismatch is structural. An attacker who clears authentication, whether through stolen credentials or a deepfaked voice, can escalate privileges and reach Active Directory before manual verification has a chance to intervene.
Yet most enterprises still run their highest-value identity decisions through manual checkpoints. Help desk password resets are authenticated by voice. Financial approvals confirmed over video. Vendor onboarding validated by document review. Every one of those workflows assumes the human on the other end is real. That assumption has become the primary attack vector, and no amount of security awareness training patches it when attackers automate at machine speed.
The imbalance is a choice, not a constraint. Closing the gap requires embedding identity verification into the systems themselves, not bolting it on as a human checkpoint.
Some organizations are moving past manual workflows entirely. Carter Rees, vice president of Artificial Intelligence at Reputation, told VentureBeat that identity must become a native input to AI reasoning itself. “We are moving toward an identity-embedding framework where role-based permissions and behavioral baselines are encoded directly into model reasoning, not just enforced in admin dashboards,” Rees said.
Active campaigns targeting enterprise identity
“We’re seeing threat actors use GenAI to scale social engineering, accelerate operations, and lower the barrier to entry for hands-on-keyboard intrusions,” Adam Meyers, Senior Vice President of Counter Adversary Operations at CrowdStrike, told VentureBeat.
FAMOUS CHOLLIMA, a DPRK-nexus adversary tracked by CrowdStrike, used GenAI to create synthetic identities, deepfake interview personas, and AI-assisted coding to infiltrate over 320 companies. That represents a 220% year-on-year increase according to the CrowdStrike 2026 Global Threat Report.
These are not theoretical attack chains. They are running operations with documented enterprise victims.
The identity-layer risk extends beyond authentication. Rees flagged an emerging attack surface that most security teams have not yet mapped. “User embeddings are sensitive identity artifacts,” he told VentureBeat. “They can expose PHI or PII through inversion attacks or bias if not controlled. Security leaders must treat embeddings like credentials.” As organizations encode identity into AI systems, those encodings become targets. The same governance applied to passwords and tokens needs to extend to identity representations inside models.
Seven controls for the next 30 days
Map every workflow where a human verifies another human. Help desk password resets by voice. Financial approvals over video. Executive authorization by callback. List them all. Every one is a deepfake attack surface. Assume all are compromised until re-architected.
Deploy phishing-resistant MFA across all privileged accounts within 30 days. FIDO2 and WebAuthn bind authentication to physical hardware. Synthetic media alone cannot forge a hardware attestation challenge. Push notifications, SMS codes, and voice-based verification can all be intercepted or spoofed. Hardware-bound authentication remains the strongest MFA class available against deepfake-assisted social engineering.
Implement deepfake detection on your highest-value workflows. Audio liveness detection, video artifact analysis, and behavioral biometrics can flag synthetic media before it reaches a decision-maker. These tools carry meaningful false-positive rates and sophisticated attackers can circumvent first-generation models, so treat detection as a tripwire, not a guarantee. For the hundreds of voice and video workflows that will take months to migrate to FIDO2, detection is the interim control that narrows the gap.
Require out-of-band verification for any financial transaction above a defined threshold. When a single channel, whether video, voice, or email, serves as the sole point of trust, attackers only need to compromise that one channel. A second channel, a verified callback to a known number or a cryptographic confirmation via a separate device, breaks the attack chain.
Enforce a 72-hour patch cycle for critical identity-adjacent systems. Threat actors are reverse-engineering patches within that window. Authentication servers, SSO infrastructure, directory services, federation endpoints: prioritize these over lower-risk assets. Every hour past 72 on a critical identity system is an exposure.
Reclassify deepfake risk from “fraud” to “identity security” in the budget. If deepfake defense sits with the anti-fraud team and identity infrastructure sits with IAM, nobody owns the gap between them. Consolidate ownership. Assign a single leader accountability for synthetic identity risk across onboarding, authentication, and transaction verification.
Audit AI agent identities alongside human ones. Machine identities already outnumber human users across the average enterprise. Every AI agent with persistent system access needs the same identity governance as a privileged human account: rotation, behavioral baselining, least-privilege scoping, and anomaly detection.