The wildcard was hiding in plain sight.
Not in the camera firmware. Not in a zero-click exploit chain. In a message broker, exposed to anyone willing to register an account. CVE-2026-33356 is a broken access control flaw in the MQTT broker that powers Meari Technology's entire smart home stack. The broker does not enforce per-device subscribe access control lists. Any authenticated CloudEdge account can issue a single subscribe call to the wildcard topic 'meari/#' and receive live data from every device on the platform. Not one camera. All of them.
Researchers captured 14,204 MQTT messages from 2,117 distinct devices in five minutes.
Let that number settle. That is not a breach. It is a structural property of the architecture. The system was never designed to keep one device's data out of another account's hands. The closest analogy in blockchain terms: a shared sequencer with no state isolation, where every rollup reads every other rollup's memory.
This looks like an IoT story. It is not. It is the same story the crypto market keeps buying into Layer 2s: a proliferation of branded wrappers around one shared, weak foundation. 300 white-label brands. One codebase. One broker. 1.1 million cameras across 118 countries. The market is already pricing the next narrative before the last one's code was audited.
Where the code forks, we find the fold.
Context
Meari Technology is a Hangzhou-based original device manufacturer. ODM is a dirty word that deserves more attention in crypto: a company that builds the entire technical substrate — firmware, cloud backend, mobile applications — and sells it to brands that put their logo on the box. Meari's customers span more than 300 white-label brands. Arenti, BOIFUN, COCOCAM, PetTec, SV3C, Joystek, Luvion, Vimar. If you own any of those cameras, you are not running the brand's software. You are running Meari's.
The financial profile is instructive. The global baby monitor market is worth $1.87 billion, and the United States accounts for roughly $540 million of that. Meari went public on the Shenzhen ChiNext board in March 2025. The share price doubled in two trading days.
That chart tells you everything about how the market allocates trust. Investors bought the smart home growth narrative. Nobody audited the broker. The same dynamic that doubles a crypto token on a whitepaper applied to a consumer hardware supply chain: narrative first, code never.
I approach this with a particular bias. In 2017, during my final year of software engineering, I audited the Ethereum Classic codebase ahead of the DAO-style fork and found an integer overflow in the EVM implementation that could have drained millions during the transition. I patched it four hours before the network split. That experience rewired how I evaluate every system, whether it calls itself a Layer 1 or a baby monitor.
The whitepaper is a map. The code is the terrain. The box is marketing. Meari's response to the disclosure — coordinated over a 70-day window by Tod Beardsley of runZero — was to label the affected products 'obsolete.' Meanwhile, the MQTT broker kept streaming. There has been no confirmed GDPR Article 34 direct user notification. No one told the people whose nurseries are on camera. They found out, if they found out at all, from the security research community.
A security response is a governance decision. And here, the decision was cheap: disavow the hardware, protect the IPO, hope the regulators remain distracted. This is not incompetence. It is a deliberate allocation of liability. The ledger remembers what the market forgets, and right now it is remembering on a schedule that no retail buyer signed up for.
Core
MQTT stands for Message Queuing Telemetry Transport. It is a publish-subscribe protocol that has become the default nervous system for IoT. Devices publish to topics. Clients subscribe to topics. A broker routes the traffic. The protocol is over two decades old, and its security model depends entirely on the person who configures the broker.
In any correctly configured system, the broker holds an access control list. Each device is assigned a unique topic namespace — think 'devices/ID/status' — and each authenticated principal is authorized to subscribe only to its own namespace. A client that attempts to subscribe outside its namespace receives an authorization failure. The protocol is explicit about this.
CVE-2026-33356 says Meari's broker skips that step. The research, presented as '1.1 Million Cameras, One Wildcard: Architectural Surveillance in an IoT Cloud,' demonstrated that any authenticated CloudEdge account can subscribe to a wildcard covering every device topic on the platform. The consequences are not theoretical. The authors captured 14,204 MQTT messages from 2,117 distinct devices in five minutes.
Let's do the arithmetic, because the rate matters. Divide 14,204 messages by 2,117 devices and you get roughly 6.7 messages per device in a 300-second window. That is approximately one event every 45 seconds per camera. This is not a desultory heartbeat. That cadence represents a home in active use: motion triggers, person detection, audio events, door and window sensors, camera power cycling, firmware telemetry. An attacker sitting on that topic stream does not need to watch video to know when a family is asleep, when they are at work, when the children are in the nursery. The message graph is the floor plan. The timestamps are the schedule. The metadata alone is a surveillance product; whatever the payloads contain is gravy.
MQTT has a feature called retained messages. A retained message is the last published message on a topic, stored by the broker and delivered immediately to any new subscriber. This is designed for state synchronization — a camera tells a newly connected hub its current status without waiting for the next event. But under a wildcard subscription, retained messages deliver something far more dangerous: an instant, up-to-the-second state map of every device that has published recently. The attacker does not wait for the stream to flow. The stream is already waiting.
Draw the comparison to DeFi. This is the same architecture as a shared sequencer with no transaction isolation. Every user believes they are transacting on their own chain. In fact, every user is writing to the same mempool with everyone else's read access. The ledger layer does not enforce separation of concerns. The entire platform rests on the assumption that the operator will never misbehave, and the operator here is a single ODM with a public market valuation and a marketing department.
The second-order effect is where my own work comes in. I spent the last two years building a protocol where autonomous AI agents settle bets on-chain using options, and I personally audited the smart contracts governing the agent collateralization logic. My non-negotiable was verifiable execution: even if the AI model produces garbage decisions, the financial settlement layer remains finite, collateralized, and immutable. That floor can hold. The accounting layer can be trustless.
But agents have graduated from text to perception. Alexa, Google Home, and Home Assistant now take camera feeds as first-class inputs. Your agent sees through the camera. The moment a camera feed enters the agent's perception stream, the agent inherits that camera's security model. And if that model is a wildcard broker in Hangzhou, then the agent's 'eyes' are not autonomous. They are a remote operator's surveillance feed wearing a friendly assistant's face.
This is the oracle problem made visceral. The smart contract can guarantee that money moves only when conditions are met. It cannot guarantee that images are not being copied. The settlement layer can be verified. The perception layer is a trust assumption with a brand logo on it. And we are wiring this unverified perception layer into agents that will, in the next market cycle, have the authority to move money based on what they see.
My critique of Layer 2s applies here verbatim. The ecosystem has shipped dozens of L2 chains that slice already-scarce liquidity into fragments while calling it scaling. It is not scaling. It is the same small user base, subdivided by an increasingly complex value chain, paying more cross-domain rent for the privilege of believing they have progressed. Meari is the same disease with a different vector: three hundred brands is not a market. It is 1.1 million devices chopped into three hundred slices by a single ODM. Each brand contributes to the aggregate attack surface while claiming plausible deniability. The brands take the revenue. The broker takes the data. The consumer takes the wildcard.
The structural fragility is identical. Floor cracks reveal the foundation's weight. Meari just showed us the basement, and the foundation was a single broker with no access control list. One shared MQTT broker with a missing ACL is a foundation that fails all floors at once. That is the brutal geometry of shared infrastructure, and it is exactly why the 'many brands, real competition' story is comedy. There is no competition when everyone rents the same cloud from the same factory.
Now the disclosure record. A 70-day window is a serious coordinated disclosure timeline. It indicates the researcher behaved professionally, gave the vendor room to remediate, and published only after the window elapsed. What did the vendor do with that time? Meari's initial response was to classify the affected products as obsolete. Declaring a product obsolete does not remove it from the 1.1 million device population. It does not disconnect the broker. It does not notify the affected users. It externalizes the liability to the users while retaining the data processing capability. In my audit experience, this pattern has a name: liability arbitrage. The vendor extracts the economic value of the installed base and transfers the legal and reputational risk to the owner.
Think about this in options vocabulary. A wildcard subscription is an unlimited short position on every private moment the camera can capture. The risk is not the position you hold. It is the forward volatility of the exposure, and the consumer cannot hedge it. No strike price, no expiration, no counterparty. The only mitigation is unplugging the device, which most users will never do.
The GDPR analysis is blunt. Article 33 requires a data controller to notify the supervisory authority within 72 hours of becoming aware of a breach. Article 34 requires notification to affected individuals where the breach is likely to result in a high risk to their rights and freedoms. A wildcard subscription channel that exposed live message traffic from over two thousand devices in a demonstration window — and was available to every authenticated account on the platform — is squarely within that definition. No confirmed Article 34 notification has been made public. The people whose bedrooms are on the platform were left to discover their exposure through a DEF CON presentation.
I will not moralize about China's Personal Information Protection Law or the absence of a U.S. federal privacy statute beyond noting the asymmetry: the largest single market for these devices, the United States, has the weakest statutory enforcement. This is why the Hong Kong versus Singapore regulatory competition matters to crypto. Jurisdictions that treat user protection as a selling point will attract the institutional flows that lax-enforcement jurisdictions will eventually bleed. Meari is a case study in what happens when a market rewards narrative and ignores architecture: an IPO that doubled in two days, followed by a structural vulnerability that no financial statement will ever quantify. Governance is not a vote; it is a vector. This vector points at a statute book that nobody is enforcing.
Contrarian
The decentralized-maximalist response to Meari is predictable: put the cameras on-chain. DePIN networks, tokenized mesh infrastructure, homomorphic encryption, verifiable enclaves. Sell the narrative that cryptographic decentralization solves the surveillance architecture problem.
It does not. It amplifies it. Decentralizing an insecure trust model is like multi-sigging a vault door that was never locked. The wildcard does not disappear. It is replicated across a blockchain's global mempool. The cryptographic guarantee can certify provenance of the video, but it cannot prevent the sensor from being a surveillance endpoint in the first place. The layer you decentralize matters. The settlement layer can be made trustless. The perception layer is still a CMOS sensor attached to someone else's cloud. No token model changes the physics.
I have spent a decade believing that code is the ultimate truth. That conviction survives. But it has been refined into a narrower claim: code is truth only where code has actually been executed and verified. The MQTT broker at issue here is code that is running. Its behavior is observable. The market's pricing of Meari's stock, like the pricing of most L2 tokens, is a narrative bet on what the code might someday do, not what it is doing today.
The closest prior art is the Zbtlink ENDLESSDOORS vulnerability, which targeted a backdoor in the infrastructure layer of embedded devices. This is different. There is no backdoor. The door was specified in the architecture. An open door is not a breach; it is a blueprint. The vulnerability class matters less than the structural thesis: the devices we trust our families to came from a commercial ecosystem where security is a cost center, not a covenant. The threshold of effort for the attacker was a single API call, and the only thing between the attacker and the nursery was a marketing department's claim.
The contrarian trade is not a decentralized hardware token. It is a short on the liability. The baby monitor market is worth nearly two billion dollars, and the U.S. segment alone over half a billion. The exposure — live feeds of homes, nurseries, and private spaces — has catastrophic downside that current prices do not reflect. When the first real-world exploitation lands, whether an extortion campaign built on a wildcard feed or a civil suit that forces discovery, the repricing will be brutal. The market will look at the 1.1 million device count and realize it is a 1.1 million victim count. That repricing event has not been optioned. The fear is unpriced. Hedging is the art of profiting from fear. The fear is there; the market just has not priced it.
The winning consumer behavior, and the winning investment behavior, is boring. Local-first, self-hosted, open-sourced automation: Home Assistant treats privacy as a competitive moat, not a default assumption. No wildcard subscription reaches it because there is no cloud account with global read access. This is the same discipline I applied when I built arbitrage bots during the Yuga Labs floor crash: find the mispriced mechanic, extract value from the market's emotional recoil, and do it with code, not conviction. The local-first architecture is the code. The conviction is expensive.
Takeaway
Your AI agent will execute on-chain. It will be collateralized, verified, and immutable — provided the code holds. But it will see through a camera's eyes. And if those eyes belong to the Meari wildcard, your trustless agent is a tenant in a house with a lock designed to be open by design.
The ledger remembers what the market forgets: 1.1 million cameras, one wildcard, and a stock that doubled while the broker streamed. Next time the market prices a narrative, ask who owns the code.
The fix was never a better camera. The fix is a broker that refuses a wildcard, a regulator that enforces Article 34, and a market that discounts convenience by the cost of exposure.
The code forks when the foundation cannot hold. Where you choose to build the floor for your own home — that decision belongs to you.
The cameras are watching.