Think Like An Attacker? Flip that advice!

For many years, I have been saying that “think like an attacker” is bad advice for most people. For example:

Here’s what’s wrong with think like an attacker: most people have no clue how to do it. They don’t know what matters to an attacker. They don’t know how an attacker spends their day. They don’t know how an attacker approaches a problem. Telling people to think like an attacker isn’t prescriptive or clear.

And I’ve been challenging people to think like a professional chef to help them understand why it’s not useful advice. But now, I’ve been one-upped, and, depending on audience, I have a new line to use.

Last week, on Veracode’s blog, Pete Chestna provides the perfect flip of “think like an attacker” to re-frame problems for security people. It’s “think like a developer.” If you, oh great security guru, cannot think like a developer, for heavens sake, stop asking developers to think like attackers.

RSA: Time for some cryptographic dogfood

One of the most effective ways to improve your software is to use it early and often.  This used to be called eating your own dogfood, which is far more evocative than the alternatives. The key is that you use the software you’re building. If it doesn’t taste good to you, it’s probably not customer-ready.  And so this week at RSA, I think more people should be eating the security community’s cryptographic dogfood.

As I evangelize the use of crypto to meet up at RSA, I’ve encountered many problems, such as choice of tool, availability of tool across a set of mobile platforms, cost of entry, etc.  Each of these is predictable, but with dogfooding — forcing myself to ask everyone why they want to use an easily wiretapped protocol — the issues stand out, and the companies that will be successful will start thinking about ways to overcome them.

So this week, as you prep for RSA, spend a few minutes to get some encrypted communications tool. The worst that can happen is you’re no more secure than you were before you read this post.

What to do for randomness today?

In light of recent news, such as “FreeBSD washing Intel-chip randomness” and “alleged NSA-RSA scheming,” what advice should we give engineers who want to use randomness in their designs?


My advice for software engineers building things used to be to rely on the OS to get it right. That defers the problem to a small number of smart people. Is that still the right advice, despite recent news? The right advice is pretty clearly not that a normal software engineer building in Ruby on Rails or asp.net should go and roll their own. It also cannot be that they spend days wading through debates. Experts ought to be providing guidance on what to do.

Is the right thing to hash together the OS and something else? If so, precisely what something else?

The Psychology of Password Managers

As I think more about the way people are likely to use a password manager, I think there’s real problems with the way master passwords are set up. As I write this, I’m deeply aware that I’m risking going into a space of “it’s logical that” without proper evidence.

Let’s start from the way most people will likely come to a password manager. They’ll be in an exploratory mood, and while they may select a good password, they may also select a simple one that’s easy to remember. That password, initially, will not be protecting very much, and so people may be tempted to pick one that’s ‘appropriate’ for what’s being protected.

Over time, the danger is that they will not think to update that password and improve it, but their trust in the password manager will increase. As their trust increases, the number of passwords that they’re protecting with a weak master password may also increase.

Now we get to changing the master password. Assuming that people can find it, how often will someone wake up and say “hey, I should change my master password?” Changing a master password is also scary. Now that I’ve accumulated hundreds of passwords, what happens if I forget my new password? (As it turns out, 1Password makes weekly backups of my password file, but I wasn’t aware of that. Also, what happens to the old files if I change my master password? Am I now exposed for both? That’s ok in the case that I’m changing out of caution, less ok if I’m changing because I think my master was exposed.)

Perhaps there’s room for two features here: first, that on password change, people could choose to have either master password unlock things. (Encrypt the master key with keys derived from both the old & new masters. This is no less secure than having backups available, and may address a key element of psychological acceptability.) You’d have to communicate that this will work, and let people choose. User testing that text would be fascinating.

A second feature might be to let people know how long they’ve been using the same master password, and gently encourage them to change it. This one is tricky mostly because I have no idea if it’s a good idea. Should you pick one super-strong master and use it for decades? Is there value to changing it now and again? Where could we seek evidence with which to test our instincts? What happens to long term memory as people age? Does muscle memory cause people to revert their passwords? (I know I’ve done it.) We could use a pattern like the gold bar to unobtrusively prompt.

A last element that might improve the way people use master passwords would be better browser integration. Having just gone to check, I was surprised how many sites my browser is tracking. Almost all of them were low value, and all of them now are. But why do we have two places that can store this, especially when one is less secure than the other. A browser API that allows a password manager to say “I’ve got this one” would be a welcome improvement.

Studying these ideas and seeing which ones are invalidated by data gathering would be cool. Talking to people about how they use their password managers would also be interesting work. As Bonneau has show, the quest to replace passwords is going to be arduous. Learning how to better live with what we have seems useful.

1Password & Hashcat

The folks at Hashcat have some interesting observations about 1Password. The folks at 1Password have a response, and I think there’s all sorts of fascinating lessons here.

The crypto conversations are interesting, but at the end of the day, a lot of security is unavoidably contributed by the master password strength. I’d like to offer up a simple contribution. Agilebits should make two non-cryptographic changes in addition to any crypto changes.

These relate to the human end of the issue, and how real humans make decisions. That is, picking a master password is a one time event, and even if there’s a strength meter, factors of memorability, typability, etc all come into play when the user selects a password when first installing 1Password.

Those human factors are not good for security, but I think they’re addressable.

First, the master password entry screens should display the same password strength meter that’s displayed everywhere else. It’s all well and good to discuss in a blog post that people need strong master passwords, but the software should give regular feedback about the strength of that master password. Displaying a strength meter each time it’s entered creates some small risk of information disclosure via shoulder-surfing, and adds pressure to make it stronger.

Second, they should make it easier to change the master password. I looked around, couldn’t figure out how to do so in a few minutes. [Update: It’s in preferences, security. I thought I’d looked there, may have missed it.]

1password

If master passwords are so important, then it’s important for the software to help its customers get them right.

There’s an interesting link here to “Why Johnny Can’t Encrypt.” In that 1999 paper, Whitten and Tygar made the point that all the great crypto in PGP couldn’t protect its users if they didn’t make the right decisions, and making those decisions is hard.

In this case, the security of password vaults depends not only on the crypto, but also on the user interface. Figuring out the mental models that people have around password storage tools, and how the interface choices those tools make develop those mental models is an important area, and deserves lots of careful attention.

Does 1Password Store Passwords Securely?

In ““Secure Password Managers” and “Military-Grade Encryption” on Smartphones: Oh, Really?” Andrey Belenko and Dmitry Sklyarov write quite a bit about a lot of password management tools. This is admirable work, and I’m glad BlackHat provided a forum for it. However, as a user of 1Password, I was concerned to read the following about that program:

However, because PKCS7 padding is used when encrypting database encryption key, it is possible to verify password just by computing KEK (using MD5 hash function), decrypting last block of encrypted database key, and checking if it equals to 16 bytes with value 0x10 (this will be the PKCS7-compliant padding when encrypting data whose length is exactly N blocks of underlying cipher). Thus, very fast password recovery attack is possible, requiring one MD5 computation and one AES trial decryption per password.

As a result of this design issue, password guessing against passwords [stored by 1Password for iPhone] is estimated (by Belenko and Sklyarov) as 15 Million per second. This is the 3rd worst performance out of a group of 11, and 3,000-fold worse than the best performer in the table (Strip Lite Password Manager, at 5,000 per second).

The folks at Agile Bits, makers of 1Password took the time to blog about the paper, and accept the implications of the work in “Strong Security Requires Strong Passwords.”

However, I think they misunderstand the paper and the issue when they write:

The main reason the password can be determined so quickly is because 6 characters provide relatively few possible password combinations.

I believe the main reason for the issue is because of the way in which 1Password has chosen to store passwords. They alude to this further down in the post when they write:

With that said, as Dmitry and Andrey point out, 1Password could do more to slow the password discovery process, thereby making it take even longer. For example, on the desktop (both Windows and Mac), 1Password uses PBKDF2 to significantly slow down attackers. Currently this is not available on iOS as we needed to support older devices. The next major release of 1Password will only support iOS 5 and at that time we will be incorporating these additional defences.

I still don’t think that’s an adequate response. Several of their competitors on iOS use their own implementation of PBKDF2. Now that’s a risky thing to do, and I’m aware that it might be expensive to implement and test, and the impact of a bug in such code might reasonably be pretty high. So it’s not a slam dunk to do so, in the general case. But in this case, it appears that Apple ships an open source version of PBKDF2: http://opensource.apple.com/source/CommonCrypto/CommonCrypto-55010/Source/API/CommonKeyDerivation.c. So the risk is far lower than creating a new implementation. Therefore, I think Agile Bits should change the way it validates passwords, and incorporate PBKDF2 into all versions of 1Password soon.

They also state:

1Password for iPhone will no longer allow items to be protected by just the PIN code. The PIN code was meant for less sensitive items and we always expected the Master Password protection to be enabled on important items. To simplify things, all items will be protected with the Master Password, just like on iPad, Mac, and Windows.

I understand the choice to do this, and move to stronger protection for all items. At the same time, I like the PIN-only protection for my low-value password. Entering passwords on a phone is a pain. It’s not an easy trade-off, and a 4-digit PIN is always going to be easy to brute force with modern CPUs, however much salting and stretching is applied. I’m capable of making a risk management decisions, but I also understand that many people may feel that Agile Bits wouldn’t offer the choice if it wasn’t secure. I respect the choice that Agile Bits is making to force stronger protection on all their customers.

In summary, 1Password is not storing passwords as securely as they could, and if your phone is stolen, or your phone backups are accessed, those choices leave your passwords at more risk than competing products. I don’t think the fixes to this require iOS5. I think the right thing for Agile Bits to do is to ship an update with better protection against brute force attacks for all their customers, and to do so soon.

[Update 3 (April 10) Agile Bits has released an update which implements 10K PBKDF2 iterations.]

[Update 2: 1Password has now stated that they will do this, adding PBKDF2 to all versions for iOS, which had been the only platform impacted by these issues. They have a hard balance of speed versus security to make, and I encourage them to think it through and test appropriately, rather than rushing a bad fix. ]

[Updated to clarify that this applies only to the iPhone version of 1Password.]

Shocking News of the Day: Social Security Numbers Suck

The firm’s annual Banking Identity Safety Scorecard looked at the consumer-security practices of 25 large banks and credit unions. It found that far too many still rely on customers’ Social Security numbers for authentication purposes — for instance, to verify a customer’s identity when he or she wants to speak to a bank representative over the telephone or re-set a password.

All banks in the report used some version of the Social Security number as a means of authenticating the customer, Javelin found. The pervasive use of Social Security numbers was surprising, given the importance of Social Security numbers as a tool for identity theft, said Phil Blank, managing director of security, risk and fraud at Javelin. (“Banks Rely Too Heavily On Social Security Numbers, Report Finds“, Ann Carrns, New York Times)

Previously here: “Social Security Numbers are Worthless as Authenticators” (2009), or “Bad advice on SSNs” (2005).

Niels Bohr was right about predictions

There’s been much talk of predictions lately, for some reason. Since I don’t sell anything, I almost never make them, but I did offer two predictions early in 2010, during the germination phase of a project a colleague was working on. Since these sort of meet Adam’s criteria by having both numbers and dates, I figured I’d share.

With minor formatting changes, the following is from my email of April, 2010.

Prediction 1

Regulation E style accountholder liability limitation will be extended
to commercial accountholders with assets below some reasonably large
value by 12/31/2010.

Why:  ACH and wire fraud are an increasingly large, and increasingly
public, problem.  Financial institutions will accept regulation in order
to preserve confidence in on-line channel.

WRONG!

Prediction 2

An episode of "state-sponsored SSL certificate fraud/forgery" will make
the public press.

Why: There is insufficient audit of the root certs that browser vendors
innately trust, making it sufficiently easy for a motivated attacker to
"build insecurity in" by getting his untrustworthy root cert trusted by
default.  The recent Mozilla kerfuffle over CNNIC is an harbinger of
this[1].  Similarly, Chris Soghoian's recent work[2] will increase
awareness of this issue enough to result in a governmental actor who has
done it being exposed.

Right!

But only because for this one I forgot to put in a date (I meant to also say “by 12/31/2010″, which makes this one WRONG! too.

I was motivated to make this post because I once again came across Soghoian’s paper just the other day (I think he cited it in a blog post I was reading). He really nailed it. I predict he’ll do so again in 2012.

The output of a threat modeling session, or the creature from the bug lagoon

Wendy Nather has continued the twitter conversation which is now a set of blog posts. (My comments are threat modeling and risk assessment, and hers: “That’s not a bug, it’s a creature. “)

I think we agree on most things, but I sense a little semantic disconnect in some things that he says:

The only two real outputs I’ve ever seen from threat modeling are bugs and threat model documents. I’ve seen bugs work far better than documents in almost every case.

I consider the word “bug” to refer to an error or unintended functionality in the existing code, not a potential vulnerability in what is (hopefully) still a theoretical design. So if you’re doing whiteboard threat modeling, the output should be “things not to do going forward.”

As a result, you’re stuck with something to mitigate, probably by putting in extra security controls that you otherwise wouldn’t have needed. I consider this a to-do list, not a bug list.
(“That’s not a bug, it’s a creature. “, Wendy Nather)

I don’t disagree here, but want to take it one step further. I see a list of “things not to do going forward” and a “todo list” as an excellent start for a set of tests to confirm that those things happen or don’t. So you file bugs, and those bugs get tracked and triaged and ideally closed as resolved or fixed when you have a test that confirms that they ain’t happening. If you want to call this something else, that’s fine–tracking and managing bugs can be too much work. The key to me is that the “things not to do” sink in, and to to-do list gets managed in some good way.

And again, I agree with her points about probability, and her point that it’s lurking in people’s minds is an excellent one, worth repeating:

the conversation with the project manager, business executives, and developers is always, always going to be about probability, even as a subtext. Even if they don’t come out and say, “But who would want to do that?” or “Come on, we’re not a bank or anything,” they’ll be thinking it when they estimate the cost of fixing the bug or putting in the mitigations.

I simply think the more you focus threat modeling on the “what will go wrong” question, the better. Of course, there’s an element of balance: you don’t usually want to be movie plotting or worrying about Chinese spies replacing the hard drive before you worry about the lack of authentication in your network connections.

Threat Modeling and Risk Assessment

Yesterday, I got into a bit of a back and forth with Wendy Nather on threat modeling and the role of risk management, and I wanted to respond more fully.

So first, what was said:

(Wendy) As much as I love Elevation of Privilege, I don’t think any threat modeling is complete without considering probability too.
(me) Thanks! I’m not advocating against risk, but asking when. Do you evaluate bugs 2x? Once in threat model & once in bug triage?
(Wendy) Yes, because I see TM as being important in design, when the bugs haven’t been written in yet. :-)

I think Wendy and I are in agreement that threat modeling should happen early, and that probability is important. My issue is that I think issues discovered by threat modeling are, in reality, dealt with by only a few of Gunnar’s top 5 influencers.

I think there are two good reasons to consider threat modeling as an activity that produces a bug list, rather than a prioritized list. First is that bugs are a great exit point for the activity, and second, bugs are going to get triaged again anyway.

First, bugs are a great end point. An important part of my perspective on threat modeling is that it works best when there’s a clear entry and exit point, that is, when developers know when the threat modeling activity is done. (Window Snyder, who knows a thing or two about threat modeling, raised this as the first thing that needed fixing when I took my job at Microsoft to improve threat modeling.) Developers are familiar with bugs. If you end a strange activity, such as threat modeling, with a familiar one, such as filing bugs, developers feel empowered to take a next step. They know what they need to do next.

And that’s my second point: developers and development organizations triage bugs. Any good development organization has a way to deal with bugs. The only two real outputs I’ve ever seen from threat modeling are bugs and threat model documents. I’ve seen bugs work far better than documents in almost every case.

So if you expect that bugs will work better then you’re left with the important question that Wendy is raising: when do you consider probability? That’s going to happen in bug triage anyway, so why bother including it in threat modeling? You might prune the list and avoid entering silly bugs. That’s a win. But if you capture your risk assessment process and expertise within threat modeling, then what happens in bug triage? Will the security expert be in the room? Do you have a process for comparing security priority to other priorities? (At Microsoft, we use security bug bars for this, and a sample is here.)

My concern, and the reason I got into a back and forth, is I suspect that putting risk assessment into threat modeling keeps organizations from ensuring that expertise is in bug triage, and that’s risky.

(As usual, these opinions are mine, and may differ from those of my employer.)

[Updated to correct editing issues.]