The first was that 123456 was the credentials for the admin panel.
The second was an insecure direct object reference, where the lead_id querystring parameter can be changed on an API call to retrieve another applicant's data.
UUIDs aren’t “just more difficult to guess.” They are inconceivably harder to guess.
> Put another way, one would need to generate 1 billion v4 UUIDs per second for 85 years to have a 50% chance of a single collision.
Can you tell I've been scarred by discussing designs with folks who focus on the "visible" problems without thinking about the fundamental question of "is this secure"?
> The first was that 123456 was the credentials for the admin panel.
No. 123456 was the credentials for the test setup, which contained nothing. But you could use the IDOR to access data from the test setup.
If 123456 had been the credentials to the admin panel, there would have been no point in exploiting an IDOR - as an admin, you can just look at whatever you want.
With HMAC, you can still ask for some sequential IDs
SipHash128(0, KEY) = k_0
SipHash128(1, KEY) = k_1
You get the same number of bits as a UUID.
You can't, however, sort by IDs to get their insertion sequence, however. For that you'd need something like symmetric encryption but this is already a bad idea, no reason to make it worse.
The point is not that UUIDs are magically secure, it's that they mean nothing to whoever gains access except a single job app. The assumption is that they will get out (they're in a public URL), and that they will have no meaning when they do.
It's a defense-in-depth thing IMO -- cargo-culting this approach defends you even when you don't do the other things right. It's simple -- with a non-zero probability that the actual access control is faulty, do you want a default that protects you or doesn't. What's the intentional trade we're going for? More DB perf? Easier to type URLs? There are other ways to deal with those
> Can you tell I've been scarred by discussing designs with folks who focus on the "visible" problems without thinking about the fundamental question of "is this secure"?
Yes :(
def uuid4():
"""Generate a random UUID."""
return UUID(bytes=os.urandom(16), version=4)
Defense in depth is a thing, so even if you make a mistake in one place, and the attacker gets complete access - as what happened with the McApplicaton here - they won't be able to download your entire db within minutes. Even with zero authentication, non-guessable identifiers will slow down the exfiltration by several factors from dozens/hundreds of records per second to one record per $MANY_DAYS, with lots of 404s for the defenders to look at.
> That means you'll probably leak them, expose them, or other folks will collect them (often incidentally via things like system logs)
The additional friction of acquiring the UUIDs from a different channel is beneficial to defenders, compared to decrementing or incrementing IDs, which is trivial to do, and doesn't need RCE. It's the difference between "All users' data was exfiltrated" and "Only a couple/handful of accounts were affected", and this can make or break the breached company.