A new technical report from Bitsight's threat research team, covered in detail by Krebs on Security, documents one of the more architecturally sophisticated consumer IoT fraud operations to emerge in recent years. The subject is the H96, a generic Android TV streaming stick sold openly on Amazon and other major retail platforms. The finding goes well beyond the previously documented residential proxy abuse: these devices are enrolled as participants in a coordinated ad fraud network, spoofing mobile device identities to generate fraudulent clicks on AI-generated publisher websites, all orchestrated by a mainland Chinese entity called Zhejiang Fengwo IoT Technology Ltd.
The research methodology itself deserves attention before examining the findings. Bitsight researcher Pedro Falé registered an expired domain that had previously served as a telemetry endpoint for the H96 device fleet. This is a well-established intelligence-gathering technique in threat research, analogous to sinkholing command-and-control infrastructure, and it gave Falé passive visibility into tens of thousands of devices phoning home without any active intrusion. The telemetry traffic revealed device hardware profiles, installed application lists, and crucially, the spoofed device identities being broadcast by each box. That a single expired domain could yield this level of operational visibility suggests the Fengwo Group's infrastructure was not designed with operational security as a priority, which is itself an interesting signal about the threat actor's risk model.
The Phone Spoofing Mechanism and Its Operational Logic
The core technical deception is straightforward but effective. Each H96 device, physically an Android TV set-top box, presents itself to external services as a mobile handset from one of several manufacturers including Samsung, Vivo, Huawei, and Xiaomi. This is not a trivial software trick; it requires consistent modification of the device's User-Agent strings, Android device properties, and potentially the advertising identifier ecosystem to pass basic consistency checks that ad networks apply.
The reason for this impersonation is structural to how programmatic advertising works. Most premium ad inventory is allocated against mobile web traffic, and advertiser bids for mobile placements are substantially higher than for connected TV or desktop traffic. By spoofing a mobile identity, the botnet can access higher-value ad slots and generate more revenue per fraudulent click. The AI-generated publisher sites operated by the Fengwo Group add a second layer of deception: they only serve ads when the visiting client matches the expected spoofed mobile profile. This conditional ad serving is a deliberate anti-detection measure. A security researcher visiting the site from a standard browser would see nothing monetisable, making automated crawl-based detection significantly harder.
Bitsight's report notes that the Fengwo Group fuses three vision and reasoning systems into a single interface to allow bots to identify ads on a page and interact with them in a human-like manner. This is a meaningful technical claim. Combining optical character recognition, layout analysis, and some form of decision-making model to simulate browsing behaviour is consistent with what the academic literature on adversarial ad fraud describes as second-generation click fraud, where simple HTTP request replay has been replaced by genuine browser automation with perceptual awareness. The use of headless browsers on resource-constrained embedded hardware is non-trivial, which partly explains the observed behaviour of switching between proxy and fraud modes depending on television usage.
Blockly as a Fraud Automation Framework
One of the more analytically interesting findings concerns the Fengwo Group's internal tooling. The operation uses a proprietary implementation of Google's Blockly visual programming environment to define and deploy fraud routines. Blockly is a drag-and-drop code editor originally designed for educational contexts, where visual blocks correspond to underlying JavaScript constructs. In the Fengwo Group's implementation, operators compose fraud workflows by assembling blocks, which are then exported as JavaScript and uploaded to S3 buckets for distribution to the device fleet.
This is a deliberate organisational design choice with clear economic rationale. The Fengwo Group's own developer commentary, captured by Bitsight, states that only a small number of highly skilled developers are needed to build template execution-unit images, while operators creating execution units from those templates have significantly lower technical requirements. This is essentially a platform model applied to fraud infrastructure: a small engineering team maintains the core capability, while a larger, cheaper workforce operates it. The parallel to legitimate software-as-a-service platforms is uncomfortable but accurate.
From a software engineering perspective, the choice of Blockly is clever. It abstracts away the complexity of browser automation and network interaction behind a composable interface, reducing the surface area for operator error and lowering the skill floor for scaling the operation. It also means the fraud logic is modular and updatable: pushing a new Blockly module to enrolled devices is operationally equivalent to a software update, allowing the operation to adapt to detection measures or shift between fraud task types without redeploying the underlying infrastructure.
The Dual-Mode Architecture: Proxy and Fraud
The behavioural separation between residential proxy operation and ad fraud activity is architecturally significant. Bitsight found that devices switch modes based on HDMI signal detection: when a television is active and the user is streaming content, the device operates as a residential proxy node; when the television is off, it waits for and executes ad fraud jobs. The stated rationale is resource contention, since ad fraud tasks involving browser automation are computationally expensive on embedded ARM hardware and would degrade the streaming experience.
This design has a secondary benefit for the operators: it reduces the probability of user detection. A TV box that visibly stutters during streaming would prompt investigation. By confining the most resource-intensive activity to idle periods, the malicious workload remains largely invisible to the end user. It also means the device's network behaviour during active use, which is the period most likely to be observed by a technically aware user monitoring their home router, looks entirely normal.
The residential proxy component is not new. Services like Synthient have documented for years that these devices ship with proxy software pre-installed, and the FBI has issued repeated public warnings. What the Bitsight analysis adds is a clearer picture of the full economic stack: proxy revenue funds ongoing operations and provides a steady baseline income, while ad fraud represents a higher-margin overlay that activates when the device is idle. The estimated $50,000 per day in ad fraud revenue, derived conservatively from approximately 38,000 devices reporting to a single older domain, suggests the total operation across all infrastructure could be substantially larger.
Attribution, Shell Structures, and the Fengwo Group
Attribution in this case rests on several converging technical and corporate indicators. The apps found on H96 devices share SSL certificate data with the fwgcloud.com domain operated by the Fengwo Group. The Fengwo Group holds patents that match the functional description of the apps. An internal wiki platform on the Fengwo Group's domain directly references the Blockly implementation used to build fraud routines. Shell entities in Hong Kong and Singapore appear to collect monetisation revenue, with Bitsight tracing these back to the mainland Chinese parent company.
This layered shell structure is consistent with documented patterns in Chinese grey-market ad tech operations, where domestic entities establish offshore corporate vehicles to interact with Western advertising networks and payment processors. It complicates enforcement because the entities processing payments are legally distinct from the entities operating the technical infrastructure, and jurisdictional reach across Hong Kong, Singapore, and mainland China is practically limited.
The Fengwo Group's public-facing identity as a provider of 120,000 AI digital humans for companionship and customer service applications is worth treating sceptically. Bitsight notes this framing may serve as a convenient cover story, normalising the existence of a large fleet of automated agents without revealing their actual deployment. Whether or not the AI digital human product is genuine, the public framing provides plausible deniability and a legitimate-sounding business rationale for maintaining large-scale bot infrastructure.
Implications for IoT Security Policy and Consumer Practice
The H96 case illustrates a persistent failure mode in consumer IoT markets. Devices are sold through mainstream retail channels with no meaningful security vetting, running unofficial Android builds with no Play Protect certification and no path to security updates. The supply chain from Chinese ODM manufacturers to Amazon storefronts is opaque enough that malicious firmware can reach consumers without any point of accountability.
Several practical conclusions follow from this analysis:
- Consumers should verify Android TV certification status using Google's official guidance before purchasing any streaming device, and should treat any device promising unlimited content for a one-time fee as a significant risk indicator.
- Network administrators should monitor for unexpected outbound connections from set-top boxes and smart TVs, particularly to unfamiliar cloud endpoints, as a baseline hygiene measure.
- Synthient maintains a publicly accessible list of IoT devices known to ship with residential proxy software pre-installed, which is a useful reference for procurement decisions.
- The scope of the problem extends beyond streaming sticks: the FBI has documented residential proxy software in digital photo frames and other consumer IoT categories, suggesting the attack surface is broader than any single product category.
At a policy level, the continued availability of these devices on major retail platforms despite years of documented abuse points to a gap in platform liability frameworks. The devices are not illegal to sell, the malicious behaviour is not visible at point of sale, and the harm falls on advertising networks and merchants rather than directly on the purchasing consumer. This misalignment of incentives makes voluntary market correction unlikely without regulatory intervention or meaningful liability reform.
The Bitsight research also raises questions about the adequacy of current ad fraud detection. If a botnet of this scale, operating on embedded hardware with relatively unsophisticated spoofing, can sustain $50,000 per day in fraudulent revenue from a single domain cluster, the detection capabilities of programmatic advertising infrastructure are clearly insufficient. The combination of mobile identity spoofing, conditional ad serving, and human-like browser interaction represents a threat model that simple traffic analysis will not catch. The advertising industry's investment in invalid traffic detection has not kept pace with the operational sophistication of the fraud ecosystem it is trying to counter.
The Fengwo Group operation is unlikely to be unique. The economic model is proven, the tooling is deliberately designed for low-cost scaling, and the retail distribution channel remains open. What Bitsight has documented is probably one instance of a broader class of operation. The expired domain technique that enabled this research will not be available for every equivalent infrastructure cluster, which means the true scope of embedded-device ad fraud is almost certainly larger than any current estimate can capture.