OpenCode Discloses Deepfake Fraud Targeting Inference Quotas
Dax Raad, co-founder of OpenCode, revealed that someone registered for their inference service and posed as the CEO of a real company, claiming they would spend $400,000 in the coming year and requested an increase in API rate limits.
The salesperson felt something was off and searched for past videos of the CEO; the face and voice in the meeting matched almost perfectly, leading them to initially believe the client was genuine. Further investigation revealed that the individual used a domain mimicking a real company, registered only four days prior; the traffic was abnormal and originated from overseas. When the team requested additional identification, the individual claimed they were about to board a flight and could not provide materials, insisting on opening the quota first. Raad contacted the real CEO via Instagram, who was completely unaware of the situation. The CEO in the meeting was a real-time deepfake.
OpenCode, developed by Anomaly, offers coding agents and inference interfaces for developers. Raad stated he is not the CEO. In August, the team cleared 7,013 fraudulent accounts, estimating a monthly savings of about $400,000; methods included bulk registration, rotating proxies to resell tokens, and chargebacks after promotional offers, often using stolen card numbers, with traffic concentrated on six addresses at the time.
Similar attacks have expanded from transfers to computing power. In 2024, a finance staff member from Arup's Hong Kong office participated in a video meeting where the finance director and several colleagues were all fakes, ultimately leading to a transfer of about $25.6 million. Stripe CEO Patrick Collison stated that one in six newly registered users at some AI companies is a token theft; data and AI head Emily Sands noted that attackers burn through inference costs at machine speed and then disappear, with free trial abuse doubling within six months, and some clients burning over $500 in tokens for each new paid user.
Video verification is no longer the final line of defense. Buyers are groups looking to resell or use inference at low prices, while sellers operate on a quota billing model, acquiring customers through promotions. The incident was driven by large intent and speed-up requests. Beneficiaries include deepfake-as-a-service and the gray token market, while those under pressure are AI companies that prepay GPU hours to "look like customers."
Source: Public Information
ABAB AI Insight
Dax Raad transitioned from open-source coding tools to self-operated inference, turning token margins into a calculable business: graphics cards are depreciating assets, and simply powering them can yield price differences. This same price difference has made the company a target for fraud. Clearing 7,000 accounts in August was due to proxy reselling and chargebacks; this video meeting represents an upgrade in the same supply chain—first using a real face to pass the sales psychological barrier, then requesting rate limits, using agents to drain quotas within a short window. After Anomaly transitioned personal products to pay-per-use, anti-fraud shifted from payment risk control to computing power risk control.
There are two ends to the funding path. One end involves stolen cards and chargebacks, shifting cloud bills to interface providers; the other end is deepfake-as-a-service, with discussions related to the dark web increasing by about 40% year-on-year in the first five months of 2026, lowering the barrier for "fake bosses." Fraudsters are not looking for one-time wire transfers like Arup's but rather resellable inference quotas. Stripe's statistic of one in six new registrations indicates that the customer acquisition funnel has been contaminated, with network-level fingerprints like Radar becoming the new entry ticket.
Analogous situations include business email compromise evolving from emails to video meetings, and SIM swapping in the telecommunications industry: once the verification layer is breached, the asset form determines the loss form. Bank accounts correspond to wire transfers, GPUs correspond to tokens. The industry is in a phase of accelerated substitution—sales still use "video calls to see if it's really the person," while attackers have already used publicly available materials to generate real-time replicas.
Structurally, this belongs to technological substitution. The mechanism is that identity verification relies on audiovisual consistency, which can now be replicated using publicly available videos; simultaneously, tokens can be consumed at machine speed and are difficult to recover. Scarcity has shifted from "can I see the face" to "domain age, payment patterns, and a second channel for the real person." The face remains, but the responsibility is no longer on that face.
ABAB News · Laws of Cognition
Once video verification fails, quotas become cash.
A real face, real voice, real company, but the person on the other side can be fake.
Machine-speed consumption means traditional risk control cannot close the door in time.