Comment by TeMPOraL
15 hours ago
I don't buy it. 70% of the CMPs being "misconfigured" tells us that even if these panels were broken by design, the companies using them must all conveniently not notice this. Strange, given that even a small risk of large fines or prolonged legal process with public entities would warrant someone paying at least a moment of attention to this. I suspect they are, and the choice of leaving things misconfigured is deliberate.
> something added by a dev or LLM without going through proper review
FWIW, this was a problem long before LLMs were a thing, and it didn't get worse with LLMs. If anything, I'd expect LLMs to get it right by default, because ones ~everyone is using are all trained straight, they won't just silently read between the lines and write code/configs to facilitate one's illegal business model.
> I suspect they are, and the choice of leaving things misconfigured is deliberate.
That honestly doesn't fit our data or my experience. In our scanning, about 60% of the misconfigured sites had a CMP with blocking active but one or two tags bypassing consent controls
Generally those misconfigurations aren't valuable to the business. We don't see for example lots of ad targeting and conversion tracking firing without consent on an otherwise compliant site.
What we do see is things like sites with CMPs generally working, but one or two analytics events tags firing because consent wasn't properly added to a trigger, or embedded Youtube cookies set without consent, or unexpected data from a URL or query param being accidentally ingested by tracking, or devs adding performance monitoring or observability tools to applications without realising the compliance implications
There's not much business logic in paying for a CMP, blocking your own ad stack, but then letting three analytics events pass through
> FWIW, this was a problem long before LLMs were a thing, and it didn't get worse with LLMs.
This isn't supported by our experience. In the last 18 months, we've seen a big increase in ungated tracking that we catch in CI (albeit with overall much higher velocity in general). LLMs will happily add non-compliant tracking to sites, often following defaults that might be acceptable in the US but not EU. If you push back, they'll also happily implement compliant tracking, but it's definitely not the natural default you can rely on
But your comment left me curious, so I just ran an experiment via Codex -p (gpt-5.6-sol) and Opus 5 via Bedrock
Codex returned the vendor quickstart on 5 of 5 neutral prompts. It gated properly when told the company is Irish, with full Consent Mode v2 defaults denied, GA4 only mounting after consent, with a reject button
So models can produce compliant/non-compliant code based on the context you give them, which reflects what we've seen in industry
Our business is giving devs and increasingly LLMs efficient tests to check the tracking they add is as expected for the EU and then providing signed evidence packs that prove that behaviour at a given time
Another tiny data point example: A previous company used Stripe for a very niche feature (most companies who used our product didn't ever enable it) but having the SDK in the bundle drops their "anti-fraud" cookies, whatever they are). That is exactly the kind of thing that a CMP won't be able to fix without deeper engineering work, and which doesn't actually matter anyway because no one is profiting from that -- but that's the kind of thing that a company could be sued or fined for. A complete distraction from what actually matters.