Remember when the green padlock meant something? There was a stretch of years where every security guide, onboarding doc, and well-meaning IT email told users the same thing: look for the lock. Check for HTTPS. If the browser says the connection is secure, you’re fine. It was simple advice, which is exactly why it spread so well.
That advice aged badly. Recorded Future has documented darknet vendors selling counterfeit TLS certificates obtained from legitimate certificate authorities, including Comodo, Symantec, and Thawte. Prices start at $299 for the cheapest options. The vendors register these certificates using the details of real companies, which means actual business owners are walking around unaware that their corporate identity is sitting in someone’s phishing kit.
Those certificates get used for phishing operations and man-in-the-middle attacks. The padlock shows up. The browser doesn’t complain. The user has no reason to think anything is wrong, because by the only standard they were ever taught, nothing is.
Why a toolkit reviewer cares about this
I spend my time pulling apart AI tools and writing down what actually works versus what the marketing page claims. So why am I writing about certificates? Because this story is a near-perfect illustration of the thing I run into constantly when evaluating AI products: the trust signal and the trustworthiness are two separate objects, and vendors are very good at selling you the first one.
A TLS certificate is a trust signal. It was supposed to be a proxy for “this site is who it says it is.” Pay $299 on the right darknet market and you get the signal without any of the underlying property. The proxy broke, and the people relying on it are the last to find out.
Now think about how AI tools get evaluated. SOC 2 badge in the footer. “Enterprise-grade security” on the pricing page. A trust center with green checkmarks. A privacy policy that says your data isn’t used for training. Every one of those is a signal. Some of them are backed by real audits and real engineering. Some of them are backed by a design decision someone made about the footer.
I’m not claiming AI vendors are forging credentials the way these darknet sellers are. That’s a different category of behavior entirely. The point is structural: when a signal becomes the thing everyone checks, it becomes the thing worth faking, exaggerating, or technically satisfying without meaning. Certificate authorities learned this the hard way, through years of pressure over loose controls on who could buy what.
What this changes about how you vet tools
Here’s what I’ve shifted in my own review process, and what I’d suggest if you’re the person signing off on AI tools at your company.
- Treat badges as a starting question, not an answer. A compliance logo tells you a process happened. Ask which process, when, and what the scope covered. Scope is where the interesting gaps live.
- Ask where your data physically goes. Not “is it encrypted” — everything is encrypted now, including the traffic running through a man-in-the-middle setup. Ask which subprocessors touch it, which models see it, and how long it sits anywhere.
- Check whether the vendor can describe their own failure modes. Teams with solid security practices can tell you what breaks and what their response looks like. Teams with a security page can tell you they take security seriously.
- Watch for identity borrowing. The detail in this story that unsettles me most is real companies having their registration details used without their knowledge. In the AI tool space, the equivalent is logo walls. “Trusted by” sections are not sourced, not audited, and not always true.
The uncomfortable part
Browsers responded to certificate problems by changing how much weight the padlock carries and tightening what CAs can get away with. The fix came from infrastructure, not from users getting better at squinting at URL bars. That’s the realistic path for most trust problems — the people downstream can’t inspect their way out.
Which leaves AI tooling in an awkward spot, because the infrastructure layer barely exists yet. There’s no equivalent of certificate transparency logs for model providers. No standardized way to verify that a tool’s data handling matches its claims. You’re left doing what I do: reading documentation carefully, running small tests, asking annoying questions, and accepting that some portion of your assessment is trust rather than verification.
A $299 price tag on borrowed legitimacy is a useful number to keep in mind. Not because AI vendors are shopping on darknet markets, but because it quantifies something worth remembering. Signals are cheaper to acquire than substance, always. The gap between the two is where you should be spending your attention.
đź•’ Published: