The security industry’s annual Las Vegas migration — BSides LV, Black Hat USA (August 1–6), and DEF CON 34 (August 6–9, at the Las Vegas Convention Center) — landed in the same week the Senate Commerce Committee advanced four bills built on age verification.

The timing was accidental. The juxtaposition was not lost on anyone.

The argument EFF brought to Vegas

EFF sent lawyers, activists, and technologists across all three conferences, and the message on age verification was direct: these laws are ostensibly aimed at protecting children but instead put everyone at risk of security lapses, data breaches, and misuse.

The security audience is the right audience for that argument, because it is fundamentally a systems argument rather than a civil liberties one — and it survives even if you disagree entirely about the civil liberties.

Consider what an age verification mandate actually asks for as an engineering deliverable:

A new, mandatory, high-value identity database in front of every covered service. To verify age you must collect something authoritative — a government ID image, a face scan run against an age estimation model, a credit card, a carrier attestation, or a third-party identity provider check. Each of those produces a record. Records get stored, because compliance requires proof of compliance.

Aggregation at the verification layer. Because building this is expensive, most sites will not. They will integrate a third-party age verification provider. That provider then holds ID records across thousands of sites — and knows which sites each identity was verified for. That is a browsing-history-to-legal-identity mapping, concentrated at a vendor most users have never heard of and did not choose.

A permanent linkage. This is the piece that makes it categorically different from an ordinary breach. A leaked password is rotatable. A leaked credit card is reissued. A leaked government ID tied to a record of the sites you visited is not revocable. Your face and your date of birth do not change. The disclosure is permanent.

This is not hypothetical. Age verification providers have already been breached, and the resulting exposure did exactly what the model predicts.

The security framing that lands hardest with this audience: age verification does not reduce attack surface, it manufactures attack surface, and it does so by legal mandate at every covered service simultaneously. There is no threat model in which that is a net improvement.

Cindy Cohn, EFF’s executive director, spoke on the longer arc of this — drawing on her book Privacy’s Defender, covering thirty years of hackers standing against what she frames as 360-degree surveillance marshaled by both companies and governments. The talk’s premise, and the conference’s: hackers protected the internet before and can do it again.

Defeating the cameras

Bill Swearingen presented research on adversarial patterns that defeat the detection capabilities of surveillance cameras.

The underlying technique — adversarial perturbation against computer vision models — has been academically established for years. What makes this line of work matter is that modern surveillance is no longer recording; it is detecting. A camera that stores footage requires a human to review it, and human attention is the scarce resource that has always limited surveillance in practice. A camera running person detection, vehicle detection, and behaviour classification does not have that limit.

Adversarial patterns attack the classifier rather than the sensor. The camera still sees you. The model does not register you as a person.

There is an obvious asymmetry worth being honest about: these techniques are model-specific and brittle. A pattern tuned against one detector generally fails against another, and vendors retrain. This is a research result about the fragility of the systems, not a wearable countermeasure you should expect to work in the field.

But the fragility itself is the finding, and it has a policy edge. Automated surveillance systems are increasingly relied on for consequential decisions — who gets flagged, who gets stopped, who gets a citation — while resting on classifiers that a printed pattern can confuse. Systems that can be defeated by a t-shirt should probably not be generating enforcement actions without human review.

The connected devices talk nobody wants to give

DEF CON again featured research on the security and privacy of internet-connected sex toys, which is the kind of talk that gets snickered at in the abstract and stops being funny about four slides in.

The category is the purest available demonstration of IoT security economics. These devices ship with Bluetooth or WiFi connectivity, companion apps, cloud accounts, and in some cases remote control or video features. They are built by consumer product companies, not security companies. And the data they generate — usage timing, patterns, paired partners, sometimes location and video — is at the far end of the sensitivity distribution.

Prior research in this area has repeatedly found unauthenticated Bluetooth control, cloud APIs with broken access control, and app telemetry sent in the clear. The manufacturers are frequently unreachable, and there is often no update mechanism at all.

The generalisable point: the sensitivity of the data a device collects has no relationship to the security investment of the company that built it. Nothing in the market connects those two things. A company selling a $60 connected device has the same security budget whether it is a doorbell or this, and consumers have no way to evaluate the difference before buying.

Other threads from the week

  • “Privacy You Inherit: How Cultural History Writes the Source Code for AI Surveillance Policy” — on how national and cultural histories shape which surveillance practices a society finds tolerable, and how that gets encoded into AI policy.
  • Network infrastructure work on exploiting assumptions baked into commonly used network designs.
  • Browser-in-the-middle techniques targeting authenticated browser sessions — session theft after authentication, which is the same structural insight underlying this week’s passkey research: if you cannot break the auth, take what comes after it.

What to take from it

  1. If a service demands a government ID to view lawful content, understand what you are accepting. That record is permanent, it links your identity to that visit, and it will sit at a vendor you did not choose. Sometimes it is worth it. Decide deliberately rather than by reflex.

  2. Audit your connected devices for what they actually need. Anything intimate, medical, or in a bedroom: check whether it works without the cloud account. Many do. If it requires an account to function at all, that is a design decision about data, not about features.

  3. Prefer devices with no radio to devices with a radio you don’t use. The talk-worthy failures in this category are almost always in the connectivity layer, which is frequently there for a feature nobody uses.

  4. Watch the age verification litigation, not the legislation. Laws in California, Texas, and Arkansas have been substantially blocked, and those precedents are what the federal bills will run into.

The pattern

Every year, Vegas week produces a set of results that are technically interesting and politically inconvenient in the same breath. This year’s cluster was unusually coherent.

Age verification systems, automated camera surveillance, and connected consumer devices are all instances of the same thing: infrastructure deployed at scale on the assumption that it will work as designed and never be attacked. Every one of them was demonstrated to fail that assumption, in public, by people who told the vendors first.

The legislation moving through the Senate this week assumes the opposite. It assumes an identity checkpoint can be placed in front of every internet user and held securely by thousands of independent operators. There is not a person in that convention center who believes it.