45.83% of endpoints failed identity verification: how to read that industry figure

Where the number comes from and what it means

Around 2026, several industry analyses cited a figure from a CISPA, the Leibniz Center for Information Security, report: 45.83% of AI endpoints failed identity verification. The figure was then widely retold as half of all relay sites being fraudulent.

The problem with that retelling is that it collapses three different concepts into one:

1. An endpoint is not a site. One site may expose dozens of endpoints, one per model or per channel.
2. Failing verification is not the same as impersonating another model. There are many reasons verification fails.
3. A single snapshot is not a steady state. Verification is a sample, not continuous monitoring.

Why endpoints fail verification

In rough order of frequency:

1. The interface shape does not support verification

Identity verification usually relies on particular signals: a model's self-description, specific response fingerprints, and boundary behaviour around parameters. But:

• Reverse-subscription sites often do not expose standard model names and do not answer identity questions placed in a system prompt.
• Sites that rewrite prompts wrap content on the way in and out, washing the fingerprint away.
• Self-hosted gateways that map model aliases, for example mapping gpt-4o to some actual backend, defeat the method structurally.

All three fail, but their motives are entirely different: the first is a capability limit, while the other two are consequences of architecture.

2. False negatives in the method itself

Every fingerprinting method has an error rate. A model version update, an upstream switching inference clusters, or even a different default temperature can make one genuine model behave inconsistently.

People running verification rarely publish their own error rate, but readers need to know it exists.

3. Genuine model substitution

This one has to be acknowledged: selling a cheap model under an expensive model's name is a real practice in this industry, and the price gap is a strong motive, since models of the same generation can differ by five to ten times.

How to use the figure as a user

Do not use it like this

• Saying that nearly half of all sites are fraudulent, so the whole industry is untrustworthy. That scales endpoint-level data up to site level and ignores the limits of verification.
• Saying that anything outside that 45.83% is safe. Passing verification only means nothing was found at that moment and under that method.
• Using it to argue that one particular site has a problem. The figure is industry-level and has no discriminating power for a single site.

Use it like this

• Treat it as a reason you must verify for yourself, not as a conclusion. Industry-level data shows substitution is common, so you should have basic self-testing ability.
• Run a reproducible check yourself. We have a dedicated article on how to check whether a relay service is quietly downgrading your model. The core is a fixed prompt, fixed parameters and repeated sampling, comparing response characteristics. It needs no tooling and takes twenty minutes.
• Prefer sites with a complete, verifiable interface. If a site does not return /v1/models, does not accept a standard system prompt, and does not support structured output, then which model it is really running will stay a black box to you forever, and that alone should cost it points.
• Cross-check with price. If a site quotes far below comparable sites while claiming pure official resale, then at least one of cheap and pure official is inaccurate.

A more useful set of indicators

Rather than agonising over an industry-wide pass rate, watch three things you can observe continuously:

1. Uptime, meaning whether you get cut off at the worst moment.
2. Stability of response latency, not the average but the jitter at peak.
3. Consistency between the bill and the advertised rate, meaning whether you are actually charged what was published.

All three can be measured by you over time, with no third-party report involved. Our monitoring pages publish the first two, while the third is only verifiable by reconciling your own bill.

Conclusion

Industry data exists to remind you that a problem exists and how large it is, not to draw the conclusion for you. What 45.83% really says is that in this market, verification is a required step in choosing a site, not an optional one.

It is a starting point, not an answer.