“Comparing the Usability of Cryptographic APIs”

Obstacles Frame

 

(The abstract:) Potentially dangerous cryptography errors are well documented in many applications. Conventional wisdom suggests that many of these errors are caused by cryptographic Application Programming Interfaces (APIs) that are too complicated, have insecure defaults, or are poorly documented. To address this problem, researchers have created several cryptographic libraries that they claim are more usable; however, none of these libraries have been empirically evaluated for their ability to promote more secure development. This paper is the first to examine both how and why the design and resulting usability of different cryptographic libraries affects the security of code written with them, with the goal of understanding how to build effective future libraries. We conducted a controlled experiment in which 256 Python developers recruited from GitHub attempt common tasks involving symmetric and asymmetric cryptography using one of five different APIs.
We examine their resulting code for functional correctness and security, and compare their results to their self-reported sentiment about their assigned library. Our results suggest that while APIs designed for simplicity can provide security
benefits—reducing the decision space, as expected, prevents choice of insecure parameters—simplicity is not enough. Poor
documentation, missing code examples, and a lack of auxiliary features such as secure key storage, caused even participants
assigned to simplified libraries to struggle with both basic functional correctness and security. Surprisingly, the
availability of comprehensive documentation and easy-to use code examples seems to compensate for more complicated APIs in terms of functionally correct results and participant reactions; however, this did not extend to security results. We find it particularly concerning that for about 20% of functionally correct tasks, across libraries, participants believed their code was secure when it was not. Our results suggest that while new cryptographic libraries that want to promote effective security should offer a simple, convenient interface, this is not enough: they should also, and perhaps more importantly, ensure support for a broad range of common tasks and provide accessible documentation with secure, easy-to-use code examples.

It’s interesting that even when developers took care to consider usability of their APIs, usability testing revealed serious issues. But it’s not surprising. The one constant of usability testing is that people surprise you.

The paper is: “Comparing the Usability of Cryptographic APIs,” Yasemin Acar (CISPA, Saarland University), Michael Backes (CISPA, Saarland University & MPI-SWS), Sascha Fahl (CISPA, Saarland University), Simson Garfinkel (National Institute of Standards and Technology), Doowon Kim (University of Maryland), Michelle Mazurek (University of Maryland), Christian Stransky (CISPA, Saarland University), The Increasingly-misnamed Oakland Conference, 2017.

Threat Modeling Password Managers

There was a bit of a complex debate last week over 1Password. I think the best article may be Glenn Fleishman’s “AgileBits Isn’t Forcing 1Password Data to Live in the Cloud,” but also worth reading are Ken White’s “Who moved my cheese, 1Password?,” and “Why We Love 1Password Memberships,” by 1Password maker AgileBits. I’ve recommended 1Password in the past, and I’m not sure if I agree with Agilebits that “1Password memberships are… the best way to use 1Password.” This post isn’t intended to attack anyone, but to try to sort out what’s at play.

This is a complex situation, and you’ll be shocked, shocked to discover that I think a bit of threat modeling can help. Here’s my model of

what we’re working on:

Password manager

Let me walk you through this: There’s a password manager, which talks to a website. Those are in different trust boundaries, but for simplicity, I’m not drawing those boundaries. The two boundaries displayed are where the data and the “password manager.exe” live. Of course, this might not be an exe; it might be a .app, it might be Javascript. Regardless, that code lives somewhere, and where it lives is important. Similarly, the passwords are stored somewhere, and there’s a boundary around that.

What can go wrong?

If password storage is local, there is not a fat target at Agilebits. Even assuming they’re stored well (say, 10K iterations of PBKDF2), they’re more vulnerable if they’re stolen, and they’re easier to steal en masse [than] if they’re on your computer. (Someone might argue that you, as a home user, are less likely to detect an intruder than Agilebits. That might be true, but that’s a way to detect; the first question is how likely is an attacker to break in? They’ll succeed against you and they’ll succeed against Agilebits, and they’ll get a boatload more from breaking into Agilebits. This is not intended as a slam of Agilebits, it’s an outgrowth of ‘assume breach.’) I believe Agilebits has a simpler operation than Dropbox, and fewer skilled staff in security operations than Dropbox. The simpler operation probably means there are fewer usecases, plugins, partners, etc, and means Agilebits is more likely to notice some attacks. To me, this nets out as neutral. Fleishman promises to explain “how AgileBits’s approach to zero-knowledge encryption… may be less risky and less exposed in some ways than using Dropbox to sync vaults.” I literally don’t see his argument, perhaps it was lost in the complexity of writing a long article? [Update: see also Jeffrey Goldberg’s comment about how they encrypt the passwords. I think of what they’ve done as a very strong mitigation; with a probably reasonable assumption they haven’t bolluxed their key generation. See this 1Password Security Design white paper.]

To net it out: local storage is more secure. If your computer is compromised, your passwords are compromised with any architecture. If your computer is not compromised, and your passwords are nowhere else, then you’re safe. Not so if your passwords are somewhere else and that somewhere else is compromised.

The next issue is where’s the code? If the password manager executable is stored on your device, then to replace it, the attacker either needs to compromise your device, or to install new code on it. An attacker who can install new code on your computer wins, which is why secure updates matter so much. An attacker who can’t get new code onto your computer must compromise the password store, discussed above. When the code is not on your computer but on a website, then the ease of replacing it goes way up. There’s two modes of attack. Either you can break into one of the web server(s) and replace the .js files with new ones, or you can MITM a connection to the site and tamper with the data in transit. As an added bonus, either of those attacks scales. (I’ll assume that 1Password uses certificate pinning, but did not chase down where their JS is served.)

Netted out, getting code from a website each time you run is a substantial drop in security.

What should we do about it?

So this is where it gets tricky. There are usability advantages to having passwords everywhere. (Typing a 20 character random password from your phone into something else is painful.) In their blog post, Agilebits lists more usability and reliability wins, and those are not to be scoffed at. There are also important business advantages to subscription revenue, and not losing your passwords to a password manager going out of business is important.

Each 1Password user needs to make a decision about what the right tradeoff is for them. This is made complicated by family and team features. Can little Bobby move your retirement account tables to the cloud for you? Can a manager control where you store a team vault?

This decision is complicated by walls of text descriptions. I wish is that Agilebits would do a better job of crisply and cleanly laying out the choice that their customers can make, and the advantages and disadvantages of each. (I suggest a feature chart like this one as a good form, and the data should also be in each app as you set things up.) That’s not to say that Agilebits can’t continue to choose and recommend a default.

Does this help?

After years of working in these forms, I think it’s helpful as a way to break out these issues. I’m curious: does it help you? If not, where could it be better?

Secure updates: A threat model

Software updates

Post-Petya there have been a number of alarming articles on insecure update practices. The essence of these stories is that tax software, mandated by the government of Ukraine, was used to distribute the first Petya, and that this can happen elsewhere. Some of these stories are a little alarmist, with claims that unnamed “other” software has also been used in this way. Sometimes the attack is easy because updates are unsigned, other times its because they’re also sent over a channel with no security.

The right answer to these stories is to fix the damned update software before people get more scared of updating. That fear will survive long after the threat is addressed. So let me tell you, [as a software publisher] how to do secure upadtes, in a nutshell.

The goals of an update system are to:

  1. Know what updates are available
  2. Install authentic updates that haven’t been tampered with
  3. Strongly tie updates to the organization whose software is being updated. (Done right, this can also enable whitelisting software.)

Let me elaborate on those requirements. First, know what updates are available — the threat here is that an attacker stores your message “Version 3.1 is the latest revision, get it here” and sends it to a target after you’ve shipped version 3.2. Second, the attacker may try to replace your update package with a new one, possibly using your keys to sign it. If you’re using TLS for channel security, your TLS keys are only as secure as your web server, which is to say, not very. You want to have a signing key that you protect.

So that’s a basic threat model, which leads to a system like this:

  1. Update messages are signed, dated, and sequenced. The code which parses them carefully verifies the signatures on both messages, checks that the date is later than the previous message and the sequence number is higher. If and only if all are true does it…
  2. Get the software package. I like doing this over torrents. Not only does that save you money and improve availability, but it protects you against the “Oh hello there Mr. Snowden” attack. Of course, sometimes a belief that torrents have the “evil bit” set leads to blockages, and so you need a fallback. [Note this originally called the belief “foolish,” but Francois politely pointed out that that was me being foolish.]
  3. Once you have the software package, you need to check that it’s signed with the same key as before.
    Better to sign the update and the update message with a key you keep offline on a machine that has no internet connectivity.

  4. Since all of the verification can be done by software, and the signing can be done with a checklist, PGP/GPG are a fine choice. It’s standard, which means people can run additional checks outside your software, it’s been analyzed heavily by cryptographers.

What’s above follows the four-question framework for threat modeling: what are we working on? (Delivering updates securely); what can go wrong? (spoofing, tampering, denial of service); what are we going to do about it? (signatures and torrents). The remaining question is “did we do a good job?” Please help us assess that! (I wrote this quickly on a Sunday morning. Are there attacks that this design misses? Defenses that should be in place?)

Threat Modeling Encrypted Databases

Adrian Colyer has an interesting summary of a recent paper, “Why your encrypted database is not secure” in his excellent “morning paper” blog.

If we can’t offer protection against active attackers, nor against persistent passive attackers who are able to simply observe enough queries and their responses, the fallback is to focus on weaker guarantees around snapshot attackers, who can only obtain a single static observation of the compromised system (e.g., an attacker that does a one-off exfiltration). Today’s paper pokes holes in the security guarantees offered in the face of snapshots attacks too.


Many recent encrypted databases make strong claims of “provable security” against snapshot attacks. The theoretical models used to support these claims are abstractions. They are not based on analyzing the actual information revealed by a compromised database system and how it can be used to infer the plaintext data.

I take away two things: first, there’s a coalescence towards a standard academic model for database security, and it turns out to be a grounded model. (In contrast to models like the random oracle in crypto.) Second, all models are wrong, and it turns out that the model of a snapshot attacker seems…not all that useful.

IoT Security Workshop (Seattle, August)

Jean Camp and Yoshi Kohno are organizing an interesting workshop upcoming at the University of Washington on “Best Practices In The IoT:”

Our agenda begins with a presentation on the Federal Government initiatives in
the IoT. When collecting the reading materials for emerging standards, we found
nearly a thousand pages once all governmental materials are brought together…The product of the workshop will be a summary document identifying (i) a consensus set of graduated best practices for security and privacy for IoT in the home, and (ii) any gaps where best practices cannot yet be identified.

(I believe that the workshop organizers might agree with me regards the term “best practices,” but are driven by funders to use it.)

Also, they are searching for a few more sponsors if you can help in that department.

Maintaining & Updating Software

In the aftermath of Wannacry, there’s a lot of discussion of organizations not updating their systems. There are two main reasons organizations don’t update the operating systems they run: compatibility and training. Training is simpler — you have to train people about the changes to the Start Menu to move them to Windows 8, and that’s expensive. (I sometimes worked with sales people when I was at Microsoft, and they could have managed this much better than they have.)

Compatability is harder. In his excellent blog post on “Who Pays?,” Steve Bellovin discusses how “achieving a significant improvement in a product’s security generally requires a new architecture and a lot of changed code. It’s not a patch, it’s a new release.” There are substantial changes to the ways memory is managed and laid out between the versions, including ASLR, DEP, CFG, etc. There are many changes and seeing how they impact real programs is hard. That’s part of the reason Microsoft released the Enhanced Mitigation Experiment Experience Toolkit.

Rusting car

This doesn’t just apply to platforms, it also applies to libraries. (For example, see Jon Callas, “Apple and OpenSSL.”

Even when compatibility is generally very high, someone needs to test the code to see if it works, and that costs money. It costs a lot more money if you don’t have test code, test documentation (YAGNI!) or if, umm, your test code has dependencies on libraries that don’t work on the new platform…It is unlikely that re-certifying on a new platform is less than weeks of work, and for larger products, it could easily extend to person years of work, to maintain software that’s already been sold. The costs are non-trivial, which brings me back to Steve Bellovin’s post:

There are, then, four basic choices. We can demand that vendors pay, even many years after the software has shipped. We can set up some sort of insurance system, whether run by the government or by the private sector. We can pay out of general revenues. If none of those work, we’ll pay, as a society, for security failures.

This is a fair summary, and I want to add two points.

First, it remains fashionable to bash Microsoft for all the world’s security woes, there is a far bigger problem, that of open source, which is usually released without any maintenance plan. (My friends at the Core Infrastructure Initiative are working on this problem.)

  • Code is speech. The United States rarely imposes liabilities on people for speaking, and it seems downright perverse to do so more if they let others use their words.
  • There may not be an organization, or the author of the code may have explicitly disclaimed that they’re responsible. If there is, and we as a society suddenly impose unexpected costs on them, that might inhibit future funding of open source. (As an example, the author of Postfix was paid by IBM for a while. Does IBM have responsibility for Postfix, now that he’s left and gone to Google?) How does the “releasing code” calculus change if you’re required to maintain it for ever?
  • The Open Source Definition prohibits discrimination against fields of endeavor, and requires licenses be technology neutral. So it seems hard to release an open source library and forbid the use of code in long-lived consumer goods.
  • What if Bob makes a change to Alice’s code, and introduces a security bug in a subtle way? What if Alice didn’t document that the code was managing a security issue? Does she need to fix it?

Second, the costs to society will not be evenly distributed: they’re going to fall on sectors with less software acumen, and places where products are repaired more than they’re replaced, which tend to be the poorer places and countries.

[Update: Ross Anderson blogs about a new paper that he wrote with Éireann Leverett and Richard Clayton. The paper is more focused on the regulatory challenge that maintaining and updating software provokes than the economics.]

Photo by Pawel Kadysz.

The Ultimate Stopping Machine?

a metal Immobilizer around a tireSecurity is hard in the real world. There’s an interesting story on Geekwire, “BMW’s ReachNow investigating cases of cars getting stuck on Washington State Ferries.” The story:

a ReachNow customer was forced to spend four hours on the Whidbey Island ferry this weekend because his vehicle’s wheels were locked, making the vehicle immovable unless dragged. The state ferry system won’t let passengers abandon a car on the ferry because of security concerns.

BMW’s response:

We believe that the issue is related to a security feature built into the vehicles that kicks in when the car is moving but the engine is turned off and the doors are closed.

I first encountered these immobilizing devices on a friend’s expensive car in 1999 or so. The threat is thieves equipped with a towtruck. It’s not super-surprising to discover that a service like Reachnow, where “random” people can get into a car and drive it away will have tracking devices in those cars. It’s a little more surprising that there are immobilizers in them.

Note the competing definitions of security (emphasis added in both quotes above):

  • BMW is worried about theft.
  • The state ferry system is worried about car bombs.
  • Passengers might worry about being detained next to a broken car, or about bugs in the immobilization technology. What if that kicks in on the highway because “a wire gets loose”?

In “The Evolution of Secure Things,” I wrote:

It’s about the constant imperfection of products, and how engineering is a response to perceived imperfections. It’s about the chaotic real world from which progress emerges. In a sense, products are never perfected, but express tradeoffs between many pressures, like manufacturing techniques, available materials, and fashion in both superficial and deep ways.

Surprise! There’s a way to move a vehicle a long distance with the engine off, and it’s not a tow truck!

Real products, introduced into the real world, will often involve surprises like this. One characteristic of a good security architecture is that there’s the right degree of adjustability in the product, and judging that is still a matter of engineering experience.

Similarly, one of the lessons of entrepreneurship is that the problems you experience are often surprising. Investors look for flexibility in the leaders they back because they know that they’ll be surprised along the way.

Adam & Chris Wysopal webcast

(Today) Wednesday, May 24th, 2017 at 1:00 PM EDT (17:00:00 UTC), Chris Wysopal and I are doing a SANS webcast, “Choosing the Right Path to Application Security.” I’m looking forward to it, and hope you can join us!

Update: the webcast is now archived, and the white paper associated with it, “Using Cloud Deployment to Jump-Start Application Security,” is in the SANS reading room.

Certificate pinning is great in stone soup

In his “ground rules” article, Mordaxus gives us the phrase “stone soup security,” where everyone brings a little bit and throws it into the pot. I always try to follow Warren Buffet’s advice, to praise specifically and criticize in general.

So I’m not going to point to a specific talk I saw recently, in which someone talked about pen testing IoT devices, and stated, repeatedly, that the devices, and device manufacturers, should implement certificate pinning. They repeatedly discussed how easy it was to add a self-signed certificate and intercept communication, and suggested that the right way to mitigate this was certificate pinning.

They were wrong.

If I own the device and can write files to it, I can not only replace the certificate, but I can change a binary to replace a ‘Jump if Equal’ to a ‘Jump if Not Equal,’ and bypass your pinning. If you want to prevent certificate replacement by the device owner, you need a trusted platform which only loads signed binaries. (The interplay of mitigations and bypasses that gets you
there is a fine exercise if you’ve never worked through it.)

When I train people to threat model, I use this diagram to talk about the interaction between threats, mitigations, and requirements:

Threats mitigations requirements

Is it a requirement that the device protect itself from the owner? If you’re threat modeling well, you can answer this question. You work through these interplaying factors. You might start from a threat of certificate replacement and work through a set of difficult to mitigate threats, and change your requirements. You might start from a requirements question of “can we afford a trusted bootloader?” and discover that the cost is too high for the expected sales price, leading to a set of threats that you choose not to address. This goes to the core of “what’s your threat model?” Does it include the device owner?

Is it a requirement that the device protect itself from the owner? This question frustrates techies: we believe that we bought it, we should have the right to tinker with it. But we should also look at the difference between the iPhone and a PC. The iPhone is more secure. I can restore it to a reasonable state easily. That is a function of the device protecting itself from its owner. And it frustrates me that there’s a Control Center button to lock orientation, but not one to turn location on or off. But I no longer jailbreak to address that. In contrast, a PC that’s been infected with malware is hard to clean to a demonstrably good state.

Is it a requirement that the device protect itself from the owner? It’s a yes or no question. Saying yes has impact on the physical cost of goods. You need a more expensive sophisticated boot loader. You have to do a bunch of engineering work which is both straightforward and exacting. If you don’t have a requirement to protect the device from its owner, then you don’t need to pin the certificate. You can take the money you’d spend on protecting it from its owner, and spend that money on other features.

Is it a requirement that the device protect itself from the owner? Engineering teams deserve a crisp answer to this question. Without a crisp answer, security risks running them around in circles. (That crisp answer might be, “we’re building towards it in version 3.”)

Is it a requirement that the device protect itself from the owner? Sometimes when I deliver training, I’m asked if we can fudge, or otherwise avoid answering. My answer is that if security folks want to own security decisions, they must own the hard ones. Kicking them back, not making tradeoffs, not balancing with other engineering needs, all of these reduce leadership, influence, and eventually, responsibility.

Is it a requirement that the device protect itself from the owner?

Well, is it?

Threat Modeling and Star Wars

IANS members should have access today to a new faculty report I wrote, entitled “Threat Modeling in An Agile World.” Because it’s May the Fourth, I thought I’d share the opening:

As Star Wars reaches its climax, an aide approaches Grand Moff Tarkin to say, “We’ve analyzed their attack pattern, and there is a danger.” In one of the final decisions he makes, Tarkin brushes aside those concerns. Likewise, in Rogue One, we hear Galen Urso proclaim that the flaw is too subtle to be found. But that’s bunk. There’s clearly no blow-out sections or baffles around the reactor: if there’s a problem, the station is toast. A first year engineering student could catch it.

You don’t have to be building a Death Star to think about what might go wrong before you complete the project. The way we do that is by “threat modeling,” an umbrella term for anticipating and changing the problems that a system might experience. Unfortunately, a lot of the advice you’ll hear about threat modeling makes it seem a little bit like the multi-year process of building a Death Star.