I'd make that 98%. Outside of rounding errors in the margins, the remaining two percent is made up of logic bugs, configuration errors, bad defaults, and outright insecure design choices.
Disclosure: infosec for more than three decades.
OWASP top-10 was dominated by those for a very long time. They have only recently been overtaken by authorization failures.
> Note "channel updates ...bypassed client's staging controls and was rolled out to everyone regardless"
> A few IT folks who had set the CS policy to ignore latest version confirmed this was, ya, bypassed, as this was "content" update (vs. a version update)
If your content updates can break clients, they should not be able to bypass staging controls or policies.
This is going to be what most customers did not realize. I'm sure Crowdstrike assured them that content updates were completely safe "it's not a change to the software" etc.
Well they know differently now.
Also go read Parse Don’t Validate https://lexi-lambda.github.io/blog/2019/11/05/parse-don-t-va...
Looking at how this whole thing is pasted together, there's probably a regex engine in one of those sys files somewhere that was doing the "parsing"...
I agree to the second but disagree on the first. Parser generator frameworks produce a lot of code that is hard to read and understand and they don't necessarily do a better job of error handling than you would. A hand-written recursive descent parser will usually be more legible, will clearly line up with the grammar that you're supposed to be parsing, and will be easier to add better error handling to.
Once you're aware of the risks of a bad parser you're halfway there. Write a parser with proper parsing theory in mind and in a language that forces you to handle all cases. Then fuzz the program, turn bad inputs that turn up into permanent regression tests, and write your own tests with your knowledge of the inner workings of your parser in mind.
This isn't like rolling your own crypto because the alternative isn't a battle-tested open source library, it's a framework that generates a brand new library that only you will use and maintain. If you're going to end up with a bespoke library anyway, you ought to understand it well.
1. The battle testing of the generator (e.g. lex yacc is widely used)
2. The readability of the grammar file, which enables good review
You don’t (often) read the generated code, only test it. Much like you wouldn’t read the binary generated by a compiler.
1. Poorly written code in the kernel module crashed the whole OS, and kept trying to parse the corrupted files, causing a boot loop. Instead of handling the error gracefully and deleting/marking the files as corrupt.
2. Either the corrupted files slipped through internal testing, or there is no internal testing.
3. Individual settings for when to apply such updates were apparently ignored. It's unclear whether this was a glitch or standard practice. Either way I consider it a bug(it's just a matter of whether it's a software bug or a bug in their procedures).
4. This was pushed out everywhere simultaneously instead of staggered to limit any potential damage.
5. Whatever caused the corruption in the first place, which is anyone's guess.
I think most AV companies now have a helper process to do that.
If you successfully exploit the helper process, the worst damage you ought to be able to do is falsely find files to be clean.
Ought. But it depends on the way the communication with the main process is done. I wouldn't be surprised if the main process trusts the output from the parser just a tiny bit too much.
It is not possible that only this pattern caused the crash, and fuzzing omitted to try this unfuzzy pattern?
A fun example is that if you point AFL at a JPEG parser, it will eventually "learn" to produce valid JPEG files as test cases, without ever having been told what JPEG file is supposed to look like. https://lcamtuf.blogspot.com/2014/11/pulling-jpegs-out-of-th...
I'll bet you this company has some arbitrary unit test coverage requirements for PRs which developers game be mocking the heck out of dependencies. I am sure they have some vanity sonarqube integration to ensure great "code quality". This likely also went through manual QA.
However I am sure the topic of fuzz testing would not have come up once. These companies sell checkbox compliance, and they themselves develop their software the same way. Checking all the "quality engineering" boxes with very little regards for long term engineering initiatives that would provide real value.
And I am not trying to kick Crowdstrike when they are down. It's the state of any software company run by suits with myopic vision. Their engineering blogs and their codebases are poles apart.
Not staggering the updates is what blew my mind.
They probably considered it low risk, had done similar things of times hundreds of times before, etc.
Wild that anyone would consider anything in the “critical path” low risk. I would bet that they just don’t do rolling releases normally since it never caused issues before.
Not everyone is working on the same timezone.
I see remote, Israel, Canada.
https://crowdstrike.wd5.myworkdayjobs.com/en-US/crowdstrikec...
This one specifically Spain and Romania
I know they bought companies all over the globe from Denmark to other locations.
All the other engineering locations seem even less likely.
6) some form of sandboxing/error handling/api changes to make it possible to write safer kernel modules (not sure if it already exists and was just not used). It seems like the design could be better if a bad kernel module can cause a boot loop in the OS…
One of the problems with this outage was that you couldn’t even boot into safe mode without having the bit locker recovery key.
I don’t know the exact reasoning why safe mode requires the BitLocker recovery key, but presumably not doing so would open up an attack vector defeating the BitLocker protection.
Bugs in code I can almost always understand and forgive, even the ones that seem like they’d be obvious with hindsight. But this is just an egregious lack of the most basic rollout standards.
If any one of the five points above hadn’t happened, this event would have been avoided. However, if number 1 had been addressed - any of the others could have happened (or all at the same time) and it would have been fine.
I understand that we should assume that bugs will be present anywhere, which is why staggered deployments are also important. If there had been staggered deployments, the. The damage would have happened, but it would have been localized. I think security people would argue against a staged deployment though, as if it were discovered what the new definitions protected against, an exploit could be developed quickly to put those servers that aren’t in the “canary” group at risk. (At least in theory — I can’t see how staggering deployment over a 6-12 hour window would have been that risky).
There may not be out of the box fuzzers that test device drivers so you hoist all the parser code, build it into a stand-alone application, and fuzz that.
Likely this is a form of technical debt since I can understand not doing all of this day #1 when you have 5 customers but at some point as you scale up you need to change the way you look at risk.
Policy changes seem more reliable and would catch other, as of yet unknown classes of bugs.
What looks especially bad for Crowdstrike is how many things (relatively simple things) had to fail in order for this to slip through. It's like walking into Fort Knox, grabbing a gold bar, and walking out unimpeded. A complete systemic failure.
That goes equally if it was a Windows Update rolled out in one motion that broke the falcon agent/driver, or if it was Crowdstrike. There is almost no excuse for a global rollout without telemetry checks, whether it's security agent updates or os patches.
And even testing can't be trusted 100%, because writing code that does the right thing and code that tests things correctly are about equally hard, they just aren't always hard simultaneously.
what incentivized the bad decisions that led to this catastrophic failure?
It makes sense in a way - given their fast growth strategy (from nowhere to top 3) and desire to “do things differently” - the iconoclast upstarts that redefine the industry.
Or to summarise - hubris.
The "how" here is AV definition or a way to identify the attack. In CS-speak: content.
Catching 0day quickly results in good reputation that your EDR works well.
If people turn off their AV definition auto-update, they are at-risk. Why use EDR if folks don't want to stop attack quickly?
This is one bsod on windows 10. I saw another kernel panic on specific Linux distro.
What else?
One thing that is funny is that quite a few of their competitors are taking this opportunity to shit on them via Twitter and by marketing themselves as better than CrowdStrike.
Twitter, with all its issues, apparently has a feature to prevent fake news and that feature will show crowd source sentiment to debunk fake news, in this case, Twitter users showed how many times Crowdstrike competitor BSOD windows
There's best practice and then there's customer
The solution to customer reluctance to being the guinea pig is not to force every customer to be a guinea pig.
The red hat one. But they also did it to debian with a different issue, and I think another distro as well.
I'm sorry but this is the customer's fault.
If I'm using your services you work for me and you don't get to bully me into doing whatever you think needs to be done.
People that chose this solution need to be penalized, but they won't.
Compliance also has to share some of the blame here, if best practices (local testing) aren’t allowed to be followed in the name of “security”.
Many don’t have a choice, a lot of compliance is doing x to satisfy a checkbox and you don’t have a lot of flexibility in that or you may not be able to things like process credit cards which is kinda unacceptable depending on your company. (Note: I didn’t say all)
CrowdStrike automatic update happens to satisfy some of those checkboxes.
I never thought such stories were real until I encountered them…
But is there any responsibility for the clients consuming the data to have verified these updates prior to taking them in production? I haven't worn the sysadmin hat in a while now, but back when I was responsible for the upkeep of many thousands of machines, we'd never have blindly consumed updates without at least a basic smoke test in a production-adjacent UAT type environment. Core OS updates, firmware updates, third party software, whatever -- all of it would get at least some cursory smoke testing before allowing it to hit production.
On the other hand, given EDR's real-world purpose and the speed at which novel attacks propagate, there's probably a compelling argument for always taking the latest definition/signature updates as soon as they're available, even in your production environments.
I'm certainly not saying that CrowdStrike did nothing wrong here, that's clearly not the case. But if conventional wisdom says that you should kick the tires on the latest batch of OS updates from Microsoft in a test environment, maybe that same rationale should apply to EDR agents?
In the boolean sense, yes. United Airlines (for example) is ultimately responsible for their own production uptime, so any change they apply without validation is a risk vector.
In pragmatic terms, it's a bit fuzzier. Does CrowdStrike provide any practical way for customers to validate, canary-deploy, etc. changes before applying them to production? And not just changes with type=important, but all changes? From what I understand, the answer to that question is no, at least for the type=channel-update change that triggered this outage. In which case I think the blame ultimately falls almost entirely on CrowdStrike.
I used to work with regional parks and recreation departments, and they would not approve any updates that did not go through UAW environments that we had set up. All updates had to be deployed to their UAW, thoroughly tested, before going to their production environment.
I get this this is slightly different, but I'd imagine Airlines, Banks, and Hospitals would have far more strict UAW policies to avoid a single vendor from kneecapping operations.
I don't know how you could assert that this is impossible, hence channel files should be treated as code.
Honestly, it hadn't even occurred to me that software like this marketed at enterprise customers wouldn't have this kind of control already available. It seems like an obvious thing that any big organization would insist on that I just took it for granted that it existed.
Whoops.
They do, but this update bypassed all of those rules.
No one at my company apparently.
I would say on the client for buying into CrowdStrike.
And also the client for having no contingencies and just accepting a vendor pinky-swear as meaningful.
CrowdStrike failed at their responsibilities too, I just mean that so did everyone else.
When you cede your own responsibilities to someone else and don't have that backed up with contractually enforced liability to make you whole when they fuck up, and also don't provide your own contingency so it doesn't really matter what some vendor does, that's on you. That's 100% entirely on you and it doesn't matter if a million other people also did the same utterly thoughtless and lazy thing.
I understand this perspective but I think it misses the forest for the trees. You have to evaluate this kind of stuff in context. Purity tests really smack on tech message boards where nobody has any accountability to any kind of business requirements, but basically no real-world organization operates in that way, so it's all a bit irrelevant.
> When you cede your own responsibilities to someone else ...
This framing is a bit naive, I think. It isn't a boolean. Everything is about risk management, cost/benefit analysis.
My company had a lot of Azure vms impacted by this and I'm not sure who the admin was who should have tested it. Microsoft? I don't think we have anything to do with crowdstrike software on our vms. ( I think - I'm sure I'll find out this week.)
Edit: I just learned the Azure central region failure wasn't related to the larger event - and we weren't impacted by the crowd strike issue - I didn't know it was two different things. So my second part of the comment is irrelevant.
Exactly which team owns the testing is probably left up to each individual company to determine. But ultimately, if you have a team of admins supporting the production deployment of the machines that enable your business, then someone's responsible for ensuring the availability of those machines. Given how impactful this CrowdStrike incident was, maybe these kinds of third-party auto-update postures need to be reviewed and potentially brought back into the fold of admin-reviewed updates.
Which is also why you see every single customer affected - what you are suggesting is simply not an available thing to do at present for them.
At least for now - I imagine that some kind of staggered/slowed/ringed option will have to be implemented in the future if they want to retain customers.
Is there anything we can take from other professions/tradecraft/unions/legislation to ensure shops can’t skip the basic best practices we are aware of in the industry like staged rollouts? How do we set incentives to prevent this? Seriously the App Store was raking in $$ from us for years with no support for staged rollouts and no other options.
No, for exactly the reason we just saw, and the same reason why vaccines are tested before widespread rollout.
As an example, reaction time is paramount to counter many kinds of attacks - thats why blocklists are so popular, and AS blackholing is a viable option.
Agreed. It's crazy that the top tech companies enforce this in a biblical fashion, despite all sorts of pressure to ship and all that. Crowdstrike went YOLO at a global scale.
I'd assume that sort of thing would be covered in the EULA and contract -- but even if it weren't, it seems like allowing customers to define their own definition update strategy would give them a pretty compelling avenue to claim non-liability. If CrowdStrike can credibly claim "hey, we made the definitions available, you chose to wait for 2 weeks to apply them, that's on you", then it becomes much less of a concern.
This is the most interesting question to me because it doesn't seem like there is an obviously guessable answer. It seem very unlikely to me that a company like CrowdStrike pushes out updates of any kind without doing some sort of testing, but the widespread nature of the outage would also seem to suggest any sort of testing setup should have caught the issue. Unless it's somehow possible for CrowdStrike to test an update that was different than what was deployed, it's not obvious what went wrong here.
The only reason any of these things don't cause issues in practice is checksums and error correcting codes.
Errors can show up any time, and usually show up between the parts that checksums correct. On the wire, TCP protects you with a (weak) checksum. Off the wire, your computer and filesystem can still screw things up. Even CPU bugs can do this.
I've built a couple of kernel drivers over the years and what I know is that ".sys" files are to the kernel as ".dll" files are to user-space programs in that the ones with code in them run only after they are loaded and a desired function is run (assuming boilerplate initialization code is good).
I've never made a data-only .sys file, but I don't see why someone couldn't. In that case, I'd guess that no one ever checked it was correct, and the service/program that loads it didn't do any verification either -- why would it, the developers of said service/program would tend to trust their own data .sys file would be valid, never thinking they'd release a broken file or consider that files sometimes get corrupted -- another failure mode waiting to happen on some unfortunate soul's computer.
Most importantly it was never tested at all :D
Any SWE job I've worked over my entire career, nothing is deployed with new versions of dependencies without testing them against a staging environment first.
If the crowdstrike selloff continues, I'm betting this will be why.
(There's a chance I'll make trading decisions based on this rationale in the next 72 hours, though I'm not certain yet)
I've heard that said elsewhere, but I haven't found a source for it at all. Are you able to point to one for me?
Thing is, as far as I can see, deploying this database update to a Windows machine will result promptly and unconditionally in a BSOD. That implies that this update was tried on exactly zero machines before it was shipped.
The bug can't have "slipped through internal testing"; it would have failed immediately on any machine it was loaded on.
Both of these are undergraduate-level techniques. Heck, they are covered in most first-semester programming courses. Either of these failures is inexcusable in a professional product, much less one that is running with kernel-level privileges.
Bet: CrowdStrike has outsourced much of its development work.
Don’t we have those kind of failures in almost every professional product? I’ve been working in the industry for over a decade and in every single company we had those bugs. The only difference was that none of those companies were developing kernel modules or whatever. Simple saas. And no, none of the bugs were outsourced (the companies I worked for hired only locals and people in the range of +- 2h time zone)
You think everyone writing code in the US would give two shits about the quality of their output if they see the CEO pocketing another private jet while they can barley make rent?
Hell, even well paid devs in the US can be careless and lazy if their company doesn't care about quality. Have you seen some of the vulnerabilities and bugs that make it into the Android source code and on Pixel devices? And guess what, that code was written by well paid developers in the US, hired at Google l337code standards, yet would give Asian sweatshops a run for their money in terms of shittyness.
That said, I have had some experience with classic offshoring. Cultural differences make a huge difference!
My experience with "typical" programmers from India, China, et al is that they do exactly what they are told. Their boss makes the design decisions down to the last detail, and the "programmers" are little more than typists. I specifically remember one sweatshop where the boss looped continually among the desks, giving each person very specific instructions of what they were to do next. The individual programmers implemented his instructions literally, with zero thought and zero knowledge of the big picture.
Even if the boss was good enough to actually keep the big picture of a dozen simultaneous activities in his head, his non-thinking minions certainly made mistakes. I have no idea how this all got integrated and tested, and I probably don't want to know.
Sure but there's no proof yet that was the case here. That's just masive speculations based on anecdotes on your side. There's plenty of offshore devs that can run rings around western devs.
It doesn’t mean ‘Murica better, just that the origin story of staff matter, especially if you don’t have good processes around things like rca.
Every stereotype exists for a reason.
Alternatively, there might be a general assumption that lower development costs equate to inferior quality, which is a flawed yet prevalent human bias.
This problem has a promising solution, WUFFS, "a memory-safe programming language (and a standard library written in that language) for Wrangling Untrusted File Formats Safely."
HN discussion: https://news.ycombinator.com/item?id=40378433
HN discussion of Wuffs implementation of PNG parser: https://news.ycombinator.com/item?id=26714831
Ignore this failure which was catastrophic, this was a bad design asking to be exploited by criminals.
The problem wasn't storing raw memory offsets, it was not having some way to validate the data at runtime.
Yes you still need to validate the keys to avoid logic errors, but you can avoid the memory errors.
The addresses aren't being generated internal to the program, so there are no "handles". They are referencing external data by design.
That's like saying "you shouldn't use a hard-coded volatile pointer to reference a hardware device". No, you literally need to do that sometimes; especially in embedded software.
Of course it doesn't have to be either/or; you can have fast + secure, but it costs a lot more to design, develop, maintain and validate. What you can't have is a "why don't they just" simple and obvious solution that makes it cheap without making it either less secure, less performant, or both.
Given all the other mishaps in this story, it is very well possible that the software is insecure (we know that), slow and also still very expensive. There's a limit to how high you can push the triangle, but there's not bottom to how bad it can get.
What's that, three pints in a pub inside the M25? :P
Completely agree with this sentiment though, we've known that handling of binary data in memory unsafe languages has been risky for yonks. At the very least, fuzzing should've been employed here to try and detect these sorts of issues. More fundamentally though, where was their QA? These "channel files" just went out of the door without any idea as to their validity? Was there no continuous integration check to just .. ensure they parsed with the same parser as was deployed to the endpoints? And why were the channel files not deployed gradually?
For the record, the top 25 common weaknesses for 2023 are listed at:
* https://cwe.mitre.org/top25/archive/2023/2023_top25_list.htm...
Deserialization of Untrusted Data (CWE-502) was number fifteen. Number one was Out-of-bounds Write (CWE-787), Use After Free (CWE-416) was number four.
CWEs that have been in every list since they started doing this (2019):
* https://cwe.mitre.org/top25/archive/2023/2023_stubborn_weakn...
Out-of-bounds Write
Improper Neutralization of Input During Web Page Generation (‘Cross-site Scripting’)
Improper Neutralization of Special Elements used in an SQL Command (‘SQL Injection’)
Use After Free
Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection')
Improper Input Validation
Out-of-bounds Read
Improper Limitation of a Pathname to a Restricted Directory (‘Path Traversal’)
Cross-Site Request Forgery (CSRF)
NULL Pointer Dereference
Improper Authentication
Integer Overflow or Wraparound
Deserialization of Untrusted Data
Improper Restriction of Operations within Bounds of a Memory Buffer
Use of Hard-coded Credentials
Parsing is the reverse—taking an untrusted string (or binary string) that is meant to be code and converting it into a data structure.
Both are the result of taking untrusted data and assuming it'll look like what you expect, but both are not parsing issues.
Which is precisely why parsing should've been used here instead. The correct way to do this is to work at the level after parsing, not before it. "SELECT * FROM foo WHERE bar LIKE ${untrusted input}" is dumb. Parsing the query with a placeholder in it, replacing it as an abstract node in the parsed form with data, and then serializing to string if needed to be sent elsewhere, is the correct way to do it, and is immune to injection attacks.
(Probably one other factor is that SQL was designed in a peculiar way, for "readability to non-programmers", which tends to result with languages that don't map well to simple data structures. Still, there are tools that let you construct a tree, and will generate a valid SQL from that.)
HTML is a better example, because it's inherently tree-structured, and trees tend to be convenient to work with in code. There it's more obvious when you're crossing from dumb string to parsed representation, and then back.
The only situation where a parser makes sense over simple escaping routines is if you actually intended to accept a subset of the language that you're injecting into rather than plain text, in which case you'll need more than just a parser to ensure you don't have anything dangerous—you'd need to do a lot of error-prone analysis of the AST afterward as well.
You shouldn't do "escaping" and string concatenation. That's just parsing and unparsing while cutting corners, which is how you get injection bugs.
> The only situation where a parser makes sense over simple escaping routines is if you actually intended to accept a subset of the language that you're injecting into rather than plain text
And that's exactly what you're doing. With escaping, you're taking a serialized form of some data, and splice into it some other data, massaged in a way you hope will make it always parse to string when something parses this later. It's going to eventually bite you; not necessarily with XSS - web template breakage is another common occurrence.
Working in string space is tricky, dangerous, and dumb - parsing, working on the parsed representation, and unparsing at the end, is how you do it correctly and safely.
(Another way to put it: plaintext is a wire format; you don't work in it if the data is structured.)
Note that the API may look like you're doing text - see JSX - but it internally goes through a parsing stage, and makes it impossible for you to do stupid things that break or transform the program, like working in string space lets you.
If you instead escape the user-provided unstructured text by replacing the very well-known set of special characters that could create tags, you know your users cannot produce active code, only text nodes.
It's the principle of least power: if you don't need users to access anything other than unstructured text then why feed their input into a parser that produces a data structure that represents code? Make illegal states unrepresentable by just escaping the text nodes as they're saved!
Working on the data structures after parsing makes it impossible to accidentally break the structure itself. Like, maybe your string escaping is perfect, but if you do:
$content = $templatePrefix + $sanitizedString + $templateSuffix;
$result = $template.findNode("#foo").setText($unsanitizedInput)
$result = $template.findNode("#foo").setText($unsanitizedInput)
> escape the HTML using the language's standard library or the web framework's tooling.
This is what parsing the user input would look like with the DOM API:
const newDiv = document.createElement('div');
newDiv.innerHTML = untrustedUserInput;
// Do some work to attempt to sanitize the new HTML elements
document.body.appendChild(newDiv);
What you actually proposed is just escaping the HTML, not parsing user input, with the only twist being that you prefer to inject user input into your templating system imperatively with something resembling the DOM API instead of declaratively with something resembling JSX. That's fine, but not relevant to the question of what method we use to sanitize the untrusted input that we're injecting. On that front it sounds like we're in agreement that parsing user input is a terrible idea.
Surely many of these originate from deserialization of untrusted data (e.g., trusting a supplied length). It’s probably documented but I’m passively curious how they disambiguate these cases.
> Surely many of these originate from deserialization of untrusted data (e.g., trusting a supplied length).
Then they would presumably be classified under "Deserialization of Untrusted Data", number fifteen.
“Deserialization of untrusted data” isn’t even a security bug like an out of bounds write is. Every meaningful program deserializes external input. It’s a common area where bugs occur, but it’s not a type of bug in and of itself. Every bug in that category “belongs” in a more proximate category.
I wouldn't blame imperative programming.
Eg Rust is imperative, and pretty good at telling you off when you forgot a case in your switch.
By contrast the variant of Scheme I used twenty years ago was functional, but didn't have checks for covering all cases. (And Haskell's ghc didn't have that checked turned on by default a few years ago. Not sure if they changed that.)
This. One year ago UK air traffic control collapsed due to inability to properly parse "faulty" flight plan: https://news.ycombinator.com/item?id=37461695
The whole concept of "we fuck with the system in kernel based on data downloaded from the internet" is just not very sound and safe.
next time you'd be adding /s to your posts
Which is precisely the rationale which led to Standard Operating Procedures and Best Practices (much like any other Sector of business has developed).
I submit to you, respectfully, that a corporation shall never rise to a $75 Billion Market Cap without a bullet-proof adherence to such, and thus, this "event" should be properly characterized and viewed as a very suspicious anomaly, at the least
https://news.ycombinator.com/item?id=41023539 fleshes out the proper context.
28c3: The Science of Insecurity (2011)
By now, if you write any parser that deals with any outside data and don't fuzz the heck out of it, you are willfully negligent. Fuzzers are pretty easy to use, automatic and would likely catch any such problem pretty soon. So, did they fuzz and got very very unlucky or do they just like to live dangerously?
In the past ten years or so of having done somewhat serious computing and zero cybersecurity whatsoever, I have my mind concluded, feel free to disagree.
Approximately 100% of CVEs, crashes, bugs, slowdowns, and pain points of computing have to do with various forms of deserialising binary data back into machine-readable data structures. All because a) human programmers forget to account for edge cases, and b) imperative programming languages allow us to do so.
This includes everything from: decompression algorithms; font outline readers; image, video, and audio parsers; video game data parsers; XML and HTML parsers; the various certificate/signature/key parsers in OpenSSL (and derivatives); and now, this CrowdStrike content parser in its EDR program.
That wager stands, by the way, and I'm happy to up the ante by £50 to account for my second theory.