Technical explainer
Client-side scanning is not magic and not a single button marked safety. It is a sequence: content is checked at the client surface, matched or classified, routed into reporting institutions, and then handled downstream. The mechanism is documented; the risk analysis is labeled separately.
In ordinary end-to-end encryption, the service is supposed to lose access to message content once the message is protected. Client-side scanning changes the location of the check: the content is inspected on the phone or client surface before that protection fully matters. FACT
A scan can follow two paths. Known material can be compared against hash lists; unknown material can be passed through classifiers. Safer is documented as using both known-material matching and machine-learning classification, with NCMEC as a critical upstream data and reporting dependency. FACT
The risk does not require a fantasy claim. The proven technical issue is that the scanner is dual-use and expandable by policy or configuration; the analytical issue is what happens when false positives, classifier limits, and institutional dependencies are scaled across private communications. INFERENCE
Proven mechanism, documented dependency, and bounded risk are kept separate in each step.
1 / Client surface
DOCUMENTED FACT
Client-side scanning checks content on the device or client surface before the ordinary end-to-end-encryption boundary protects it.
Dependency point
Mechanism
Moving the check onto the device changes the attack surface.
This does not prove any single vendor implementation is identical.
2 / Known material
DOCUMENTED FACT
Systems such as Safer can compare content fingerprints against known-CSAM hash lists and related corpora.
Dependency point
Hash corpus
The dependency point is the corpus: who defines, maintains, and routes the matching lists.
Hash matching does not by itself prove a user's criminal intent.
3 / New material
SPLIT FACT
Machine-learning classifiers can flag new or unknown material, but documented tests and EPRS analysis keep classifier limits visible.
Dependency point
Classifier
Classifier error rates matter at platform scale.
This page does not present automated classification as reliable proof of abuse.
4 / Reporting
DOCUMENTED FACT
When a provider has actual knowledge, US-based reporting architecture routes apparent CSAM reports through NCMEC's CyberTipline.
Dependency point
Reporting chokepoint
The institutional dependency is separate from the scan itself.
A report is not a conviction and is not treated here as one.
5 / Escalation
DOCUMENTED FACT
Reports then move into human, platform, hotline, or law-enforcement review pathways; the legal status of downstream viewing remains contested in parts of US case law.
Dependency point
Institutional review
The scan is only the first institutional handoff.
The page does not collapse report volume into proven offender counts.
6 / Risk boundary
SPLIT INFERENCE
False-positive scale, classifier limits, and policy/configuration expansion are treated as documented or analytically bounded risks, while intent to exploit remains unproven.
Dependency point
Bounded risk
The dossier separates the proven security problem from unsupported claims about who would exploit it.
No claim here says US actors intend to exploit the holes.
The claim cards reuse the Technical Explainer Contract records.
DOCUMENTED FACT
This is a technical mechanism claim, not a claim that every product uses the same implementation.
DOCUMENTED FACT
The record documents the product stack and limits; it does not prove that Safer can solve scanning in encrypted spaces.
PROVEN FACT
This proves technical exploitability risk; it does not prove US actors intend to exploit the holes.
DOCUMENTED FACT
Dependency is documented; deliberate sovereignty-undermining remains an inference.
SPLIT INFERENCE
The mechanism and risk are documented; intent to exploit remains unproven.
DOCUMENTED FACT
Dependency is documented; deliberate sovereignty-undermining remains an inference.
The phone-side check creates a local decision point before the ordinary encrypted transport story begins. If the content resembles known material, a hash match can trigger the next step. If it does not, a classifier may try to predict whether the material is new CSAM or grooming-related content. The distinction matters because known-material matching and classification have different error profiles.
A hit does not end the story. It becomes a report package, enters platform and hotline handling, and can move into law-enforcement review. This page keeps that handoff visible because the dependency point is institutional as much as technical: NCMEC is a statutory US reporting chokepoint and the upstream source for parts of the hash-corpus story. FACT
A scanner that can look for one category can be repointed by policy or configuration to another category. That is the proven dual-use warning in the technical record. What this dossier adds is the bounded risk readout: once a mandate normalizes device-side inspection, the fight moves to who controls the lists, thresholds, reporting rules, and review pathways. INFERENCE
It does not prove that Thorn, NCMEC, DG HOME, or any US actor intends to exploit the holes. It does not present a report as a conviction. It does not treat a classifier score as proof of abuse. The claim strong enough to publish is narrower: the mechanism and dependencies are documented, and the risk is serious without pretending the evidence proves more than it does.
UNPROVEN
"US actors intend to exploit CSS holes" is not presented as proven here. The technical record supports exploitability and dependency risk; intent remains unproven.
SPLIT
The diagram treats a hit as the start of an escalation path, not as proof that an accused person committed an offense.