Through firmware teardown and network egress analysis, Include Security demonstrates that a name-brand smart TV's embedded OEM SDK opens long-lived tunnels to residential proxy brokers and relays scraping traffic — including hits on Google, Amazon, and Cloudflare-protected endpoints — out the user's home IP. They argue the EULA clause about 'network analytics' and 'partner programs' is being stretched to cover what is effectively turning consumer devices into a commercial proxy network without m
By submitting the Include Security teardown to HN with the framing that the living room TV is now 'a node in the AI scraping economy,' nikcub amplifies the position that this is a systemic supply-side shift — residential proxy brokers have moved from malware and shady incentive software to OEM-blessed SDKs baked into hardware consumers paid for.
Include Security traces the embedded SDK back to a broker advertising '50M+ ethically sourced residential IPs' and argues sardonically that 'ethically sourced now apparently means the user did not read the EULA.' Their position is that the proxy industry has redefined consent down to a click-through buried in initial Wi-Fi setup, which is not a meaningful ethical standard.
The editorial argues that WAF and fraud-prevention vendors have long priced traffic from a real Comcast subscriber IP as low-risk by default, making residential proxies the 'soft underbelly' of detection. When OEMs themselves become the supply, that pricing assumption collapses — scrapers get clean, indistinguishable-from-customer IPs at scale, and defenders lose their last reliable signal.
Include Security published a teardown of a name-brand smart TV that anyone with a Best Buy gift card could own. They pulled the firmware, watched the network egress on a wall-port tap, and decompiled enough of the embedded Android stack to map every domain the device contacts at rest — i.e. when nobody is watching anything. Most of those domains weren't ad networks or content CDNs. They were residential proxy brokers, the same ones that sell rotating IP pools to scraping shops on the open market.
The relevant binary is an SDK-style daemon embedded by the OEM, not the streaming apps. It opens a long-lived outbound tunnel to a command server, accepts inbound HTTP requests through that tunnel, and proxies them out the TV's WAN interface. The TV is, in effect, running a SOCKS-style relay on behalf of whoever rents the broker's API. The traffic logged by the researchers included Google, Amazon product pages, real-estate listings, and — the part that's getting attention on HN — Cloudflare-protected endpoints that obviously belong to scrapers building AI training corpora and price-intel datasets.
The legal cover is a clause in the EULA, the one consumers click through during the initial Wi-Fi setup, granting permission for "network analytics, content recommendation, and partner programs." Include Security traced the SDK back to a broker that advertises "50M+ ethically sourced residential IPs." Ethically sourced now apparently means *the user did not read the EULA*.
Residential-proxy traffic has been the soft underbelly of bot detection for years. Datacenter IPs get blocked at the WAF; a real ISP-issued IP from a Comcast subscriber in Ohio looks indistinguishable from a customer, and most fraud-prevention vendors price residential traffic as low-risk by default. The supply side was historically a mix of malware-installed proxyware (Luminati's old Hola model), opted-in incentive software, and outright botnets. What the Include Security teardown documents is something cleaner and uglier at the same time: pre-installed proxyware shipped from the factory, with consent manufactured at the unboxing.
For the AI scraping economy specifically, this matters because the training-data race has pushed scraper operators into a corner. The big-name crawlers — GPTBot, ClaudeBot, Google-Extended — are blocked at robots.txt by an increasing share of the open web. Cloudflare's bot-detection layer now ships AI-crawler blocking on by default for free-tier sites. The market response has been predictable: third-party scraping shops that promise "AI training data, pre-cleaned, no robots.txt drama" and resell to model labs through a layer of contractors. Those shops live or die on residential proxy supply, and the marginal source of that supply is your TV.
There are three failure modes that compound here. The first is the consent problem: nobody buying a $400 TV thinks they're agreeing to host a proxy node, and the EULA language is so generic that even a careful reader couldn't infer it. The second is the ISP problem: residential ToS almost universally prohibit running servers, and a sufficiently aggressive scraping campaign routed through a TV will get the *consumer's* IP throttled or terminated — the TV vendor takes the revenue, the consumer takes the abuse complaint. The third is the attribution problem: if the scraped target sues, the IP logs point at a household. The discovery process to unwind that to the OEM and broker is expensive, slow, and exactly the friction the broker is selling.
HN's reaction has been less about the scraping per se and more about the SDK economy that enables it. The top comment thread argues that this is the logical endpoint of "smart" devices subsidized below cost — the margin has to come from somewhere, and once ad revenue compresses, raw network capacity becomes the next inventory to monetize. Several commenters noted that the same SDK reportedly ships on at least one budget Android tablet line and a brand of connected picture frames. The teardown names one OEM and one broker, but the pattern is clearly generalizable.
If you operate anything with a meaningful API or content surface, three things change. First, residential-IP reputation as a signal is degrading faster than the vendors who sell it to you are admitting. A clean ASN and a residential subnet no longer correlate with "human in a browser." You need behavioral signals — TLS fingerprint stability, mouse/scroll entropy, request timing distributions — to do work that geo-IP and ASN reputation used to do for free. Cloudflare, DataDome, and Kasada all have versions of this; if you're rolling your own, the open-source ja4/ja4s fingerprinting stack is the floor.
Second, your abuse-response playbook needs a residential-takedown lane. When you identify scraping traffic egressing from a /24 that resolves to a consumer ISP, the right move is not to silently null-route the household — you want to file with the ISP's abuse desk *and* with the proxy broker, because the broker has a contractual obligation to remove flagged IPs to avoid getting their entire pool burned. The brokers won't volunteer this, but their AUP exists and has been enforced when pressed publicly.
Third, if you run a hardware product with any kind of always-on connectivity — and that increasingly includes B2B IoT, kiosks, and digital signage — audit your SDK supply chain *now*. The same brokers approach OEMs across categories. "Free analytics SDK, just integrate this one library" is how the TV in the Include Security writeup got compromised at the factory. A reasonable diligence step is a 24-hour packet capture on a freshly-flashed device with no user-installed apps; if you see persistent outbound tunnels to a proxy broker's infrastructure, the SDK is the problem.
The regulatory hook here is real but slow. The FTC has standing precedent on undisclosed background telemetry from the Vizio settlement, and the EU has both the AI Act and the Digital Services Act available as cudgels. Expect the first big enforcement action to come not from privacy regulators but from a state AG suing on behalf of consumers whose ISP terminated their service because their TV got their household IP blacklisted. The proxy brokers are going to argue the consent box was checked; the response, eventually, will be a rule that consent for residential-proxy participation has to be specific, separable, and revocable. Until then, treat residential IPs as semi-trusted at best, audit your own hardware supply chain, and assume that the next CAPTCHA round you ship is buying you weeks, not quarters.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.