SSA: deepfakes are already diverting US Social Security payments, with losses set to approach $40.1B in 2027

By Zendoric · September 26, 2026.

The Social Security Administration (SSA) has confirmed that it is coordinating with other federal agencies to curb fraud driven by deepfakes: AI-generated audio or video content that realistically mimics the voice and likeness of a real person. Chad Bungard, chief of strategy at the SSA's Office of the Inspector General (OIG), explained as much at a GovExec event in Washington: the agency's investigators have already seen cases in which these tools are used to alter direct deposit details, impersonate specific individuals and pose as trusted organizations, at times invoking the names of real employees to lend credibility to the scam.

The figure Bungard put on the table is alarming: losses from deepfake-linked fraud could reach $40.1 billion by 2027, according to Deloitte and Accenture data cited by the official himself and reported by GovCIO Media & Research. A caveat is in order: this is a projection by private consultancies repeated by a public official, not a loss already booked, and it is not broken down by agency. But the aim of the attack described —redirecting a payment already scheduled to land in an account— is concrete and verifiable enough to take seriously, above all because it strikes directly at a particularly vulnerable group: retirees and people with disabilities who depend on that monthly transfer to live.

The SSA's response, for now, is organizational more than technical. The agency is training the supervisors who work with AI systems in any capacity, and its OIG is developing governance policies and training for staff and investigators as AI gains weight both in fraud schemes and in the tools to combat them. The office already uses specialized AI for fraud detection and generative models for administrative tasks, and in August the SSA issued a request for information (RFI, a public questionnaire to gather vendor proposals) on an enterprise AI strategy that also covers so-called agentic AI: systems capable of executing multi-step tasks without constant human supervision, a field the agency itself is still working out how to approach.

On the interagency front, the OIG shares intelligence with the Justice Department's newly created National Crime Defense Center, with its own investigators and analysts embedded in its operations. It also takes part in the National Anti-Fraud Committee —whose focus is preventing fraudulent payments before they are disbursed, rather than chasing them afterwards— and in the National Fraud Detection Center, an interagency task force against large-scale fraud in federal programs. Bungard co-chairs that committee's AI subcommittee alongside the SSA's chief AI officer.

Our take: this case is almost a textbook example of the thesis we have been arguing about agentic AI and fraud: the immediate danger is not a distant superintelligence, but the cheap industrialization of scams that already existed —phone phishing, social engineering against call centers— which voice cloning and video generation make scalable at close to zero cost. The mismatch is not a lack of detection, but speed: an attacker deploys a voice clone in minutes; a federal agency takes months to publish an RFI and set up a subcommittee. That asymmetry of tempo is, in itself, the short-term vulnerability that has to be managed honestly, not an anecdote.

There is also a design pattern we have seen in other contexts —from decentralized finance to traditional banking— that applies here with total clarity: the critical point is not whether a system can recognize a fake voice, but whether a single interaction, however convincing, can on its own authorize the movement of real money. Separating the reading and preparation of a direct deposit change from its final execution —always requiring verification through an independent channel before the payment moves— matters more than perfecting the deepfake detector, which will always be one step behind the generator.

In the long run, the very move that generates this friction today —a public agency starting to operate with fraud-detection AI and exploring autonomous agents for its own administration— is also the one that could end in social benefit systems that are faster, cheaper to administer and with fraud prevented before it happens rather than chased afterwards, as the National Anti-Fraud Committee is already pursuing. The transition, however, requires investment in governance and verification to grow at the same pace as generative AI's own offensive capability, and right now it is the attacker who is winning that race.