The data shows a tool that asks you to photograph the inside of your hardware wallet. Four hundred images. A claimed accuracy above 95%. And a single disclaimer that quietly contradicts the entire pitch: a "normal" result does not guarantee the device is safe.
Start there. That last sentence is the only fully honest line in the launch β and it is the one the marketing buried.
OneKey, the hardware wallet vendor, shipped tampered.app this cycle. It is a free, no-registration web tool. You upload a photo of a circuit board. It returns a verdict on whether the device was physically tampered with. It reads OneKey's own hardware. It also reads Ledger and Trezor. The founder, Yishi, published it under his real name.
On the surface, this is a public good. Under the surface, it is a brand play wearing infrastructure clothing. Both statements are true. Separating them is the work. The tool is free, so the payment is attention and trust.
The attack surface is real, and that is why the tool sells.
A hardware wallet stores private keys offline. The threat that matters is not remote β it is physical. An adversary intercepts the device between factory and user, swaps a component, implants a chip that exfiltrates the seed phrase on first unlock. Ledger's 2023 Connect Kit compromise, the steady reports of resealed units moving through the secondary market β this is the soil the tool grows in. Genuine anxiety. Genuine attack surface. When I audited smart contract logic in 2020, the lesson was the same: security demand tracks real loss, not marketing calendars. Fear is a bad indicator, data is a leader.
So the demand is legitimate. The question is whether the instrument is.
What the tool actually is: forensics, not cryptography.
Strip the branding. The core is computer vision β an image model that compares a submitted board photo against a reference set and flags anomalies. It is post-hoc detection, not pre-hoc protection. It tells you what already happened. It does not stop what is about to happen.
That distinction matters because it sets the ceiling. Cryptographic security β a Secure Element, a signed firmware chain β prevents. Image forensics observes. One is a lock. The other is a security camera. Both have value. They are not interchangeable, and the launch framing blurs them.
Now the number. The >95% accuracy claim is the weakest link in the chain.
Here is the methodology as disclosed: roughly 400 photos, self-reported accuracy above 95%, published by the tool's own maker. No independent audit. No peer review. No third-party test.
Four hundred images is not a validation set. It is a demo. Cover three brands β OneKey, Ledger, Trezor β across multiple models and manufacturing batches, and 400 photos thin out fast. Rare tampering techniques go under-sampled. Model coverage skews. The failure modes hide in the gaps the dataset never reached.
There is a deeper problem. If the reference corpus and the test corpus overlap, the accuracy figure is circular β the model is graded on data it has effectively seen. I have no evidence this happened. I also have no evidence it did not, and neither does anyone outside OneKey. Audit the logic before you trust the label.
The privacy design cuts both ways. Not storing user photos protects the submitter and deserves recognition. It also severs the feedback loop that would let the model learn from real-world tampering β no retained corpus, no continuous improvement, unless users opt in. A security model that cannot see its own failures does not iterate. It fossilizes.
The capability boundary is undisclosed, and that is the tell.
What can the tool catch? Solder rework, obvious component swaps, gross physical anomalies β plausibly. What can it miss? Chip-level firmware implants. Logic tampering inside the Secure Element. Attacks that leave no visible trace on a board photo. The tool does not publish a "detectable versus undetectable" list. Without that list, users cannot calibrate trust. They default to over-trust. That is the danger.
Here is the contrarian read.
The value of this tool is not that it stops attacks. It raises the cost of the cheapest ones. A crude reseal job that fools a casual buyer now fails a photo check. That is real, and it is worth something. Optimize the node, secure the chain β incremental hardening compounds over time.
Watch the secondary market most closely. Second-hand hardware wallets are the highest-frequency tampering scenario β a device with unknown custody history, sold at a discount, resealed and resold. A buyer with a camera and a verdict has a cheap first filter where none existed before.
But the tool's true product is confidence, and confidence is where it gets dangerous. The biggest risk is not a missed detection. It is a false sense of security β a user who sees "normal" and stops verifying everything else. The disclaimer admits this. Credit where due: a maker who states the limit is more honest than most. But a disclaimer does not neutralize the psychology it warns about.
And note the structure. OneKey sells the wallets and grades the boards β including its competitors'. Judge and player in one. Supporting rival brands softens the conflict, but does not erase it. The tool's neutrality rests entirely on OneKey's self-restraint, with no independent governance to enforce it. Efficiency is the only honest validator, and self-reported efficiency is not a validator at all.
The takeaway is a decision framework, not a verdict.
Treat tampered.app as one input among several, never as a certificate. Buy from official channels. Verify firmware signatures on first boot. Obscure the device serial number before uploading any photo, because "we don't store them" is a policy, not a cryptographic guarantee.
Then watch the signals that will decide whether this tool earns trust: an independent security firm reproducing the accuracy claim, a public "detectable versus undetectable" list, an open-source repository that outsiders can audit, and whether Ledger or Trezor answer with a competing tool. That last one matters most. A security-tool arms race would lift the whole self-custody ecosystem.
For now, the honest position is this: a free forensics filter that raises the floor on low-grade supply chain attacks, wrapped in a brand narrative it has not yet earned. Liquidities trapped in code, not in trust β and trust is the only thing this tool is actually selling. Wait for the independent test. The red candles of a bad purchase do not negotiate with hope, but neither does a 95% claim without a second witness.
