In the last few months, I've been thinking about the problem of accidental leakage of data and how to prevent it, reverse it, and address its aftermath. This post includes various thoughts, some of them including tool-specific guidance, and some of them more general.

Examples of data leakage are as follows:

• Sending a password or sensitive credential to somebody you shouldn't send it to, or posting it in a forum with wide, even public, access.

• Sending factual information accidentally to somebody you didn't intend to share it with (or at least, didn't intend to share it with yet) or putting it in a forum with wide, even public, access.

The prevention steps are fairly similar for both cases, but the steps for addressing are somewhat different. Specifically, for the case of passwords or sensitive credentials, changing it is part of the best practice response. For factual information, on the other hand, it is often infeasible to change the facts since they're the facts!

This post is largely focused on leakage in the digital realm. Some similar ideas apply in the physical realm as well.

This post also doesn't cover large-scale data leakage, nor does it cover complementary best practices such as password security, other forms of authentication security (such as IP-based limits and two-factor authentication) and encryption. I might write additional posts about some of those topics, but those topics are in general more widely covered, hence my desire to write what I'm writing first.

I end the post with meta comments including more on its potential relevance to LessWrong as well as the distinction between general ideas and details specific to particular tools and services.

Prevention strategies: general philosophy

The accident triangle philosophy and the conjunctive nature of accidents

The accident triangle idea is that for every accident with major injury, there are several accidents with minor injury and even more accidents with no injury. A similar idea applies here: for every case of leaked sensitive data, there are several cases where something went wrong, but sensitive data didn't end up being leaked. Since sensitive data leakage is a low-probability-but-high-cost occurrence, we don't get a lot of data points for it.

One underlying insight here is that a major failure such as accidental data leakage is often conjunctive: for instance, an accidental data leakage through accidental copy/paste is a conjunction of several things:

• Failure to clear clipboard after using it for sensitive data
• Accidentally pasting it instead of whatever you wanted to paste (this might be a failure to copy something else, or accidental pressing of the paste key)
• Failure to notice this before sending

The conjunction of all these is uncommon/unlikely. But we can think of each of these as an accident-with-no-injury, and independently try to reduce the likelihood of each of them. This will reduce the overall probability of leakage due to accidental copy/paste to a negligible amount.

Avoid multitasking when handling sensitive data

In general, it's a bad idea to multitask between handling sensitive data and doing things in a completely different domain. For instance, it's a bad idea to be chatting with friends while changing a password, or composing a public tweet while editing a private doc with sensitive information. There's a possibility of getting mixed up.

Also, it's important to not do anything with sensitive data while on a conference call, especially if screensharing. It's probably also better to not work with sensitive data while in a shared physical space with people who shouldn't have access to the data, due to the risks of people seeing the data by peeking at your screen.

Develop clipboard hygiene

The failure scenario

Your computer's clipboard is the thing that gets pasted when you paste (using Ctrl+V on Windows or Cmd+V on Mac or Ctrl+Y in emacs; Linux shells vary in terms of which of these shortcuts they support). You put stuff in the clipboard by executing a cut (Ctrl+X on Windows, Cmd+X on Mac, Ctrl+W on emacs) or a copy (Ctrl+C on Windows, Cmd+C on Mac, Meta+W on emacs).

An interesting property of clipboards is that pasting doesn't clear the clipboard; the item remains in the clipboard until it is overridden by something else being cut or copied. This means that a scenario like this is not uncommon:

• I need to copy sensitive information from one place to another, so I use the clipboard for copying and pasting.

• The item remains in the clipboard.

• Later, when I'm editing a shared Google Doc or messaging a friend, I want to share a link, but I forget to execute the copy step for sharing the link, so I end up accidentally sharing the sensitive information that was in my clipboard from before.

Clipboard-clearing

• Clear the clipboard after pasting any sensitive information! One simple way to do this is to open a notepad/notes app or terminal, type a letter, then select it and copy it. Paste it to confirm that the clipboard has been updated.

• For some non-system-wide clipboards, multiple clipboard entries can coexist (this is the case with emacs, for instance). In such cases, the easiest thing I've found is to just quit and restart emacs. A friend tells me that the proper way to clear out the emacs clipboard is Meta+: followed by entering (setq kill-ring nil) for the command to eval; that does seem to work for me.

• For Linux GUIs, there is a X clipboard that stuff automatically gets into when you highlight it, even without pressing any keys to copy it (and pasting from this clipboard can be done with a right click). So if you accidentally highlight any sensitive data, it becomes part of this clipboard! A friend tells me that to clear out what's in this clipboard, it makes sense to just select a random letter or digit.

  There is a similar issue within emacs as well: when you delete text (e.g., delete a line using Ctrl+K) it enters the emacs clipboard. And since emacs keeps multiple clipboard entries, you cannot simply override it by putting something else in clipboard; you need to either quit and restart emacs, or use the command provided in the previous bullet point.

• On modern Macs, locking screen and then unlocking sleeping and then waking up from sleep clears the system-wide clipboard, so you can use this method too -- though it's generally more disruptive than just putting something else in the clipboard.

Making clipboard-clearing second nature

• After pasting data from one context to another (e.g., from your browser to your word processor, or your IDE to your terminal), ask the question: should I clear this from clipboard? And if the answer is yes, then clear it immediately.

• If you remember pasting sensitive data recently, but can't remember if you cleared it out, check what's in your clipboard (by pasting in a notepad or notes app or somewhere else that's safe). If you still see the sensitive data, that's a failure. If you see something more recent from the clipboard, that's good.

• Before starting on composing something for sharing with others, such as a Slack message, tweet, or email, or beginning to edit a shared Google Doc, check what's in your clipboard. If it has sensitive data, that's a failure.

Cutting down on dangerous clipboard use in the first place

In general, it's best not to have sensitive data in the clipboard in the first place. Consider some alternatives:

• Rather than pasting sensitive data directly, paste a link to the sensitive data, where the link is only accessible to people who have authenticated themselves.

• If you're pasting sensitive data such as a password or card number for regular entry of that password, consider other options such as using the browser autofill or a password manager.

• (A bit extreme and often not worth it) If you do have to paste sensitive data such as a password or credential, consider ways of reducing the risk of leakage as follows:

  • Paste it in two or more chunks, so that your clipboard doesn't have the whole entry at any time.

  • Paste all but the last letter, and type the last letter manually.

Cutting down on accidental pasting

The paste shortcut key (Ctrl+V in Windows, Cmd+V in Mac) is close to several other shortcut keys, including the shortcut key for copy and for bolding. So there's a good chance of accidentally pasting when you're trying to copy or bold.

If your clipboard is clear of sensitive information in the first place, accidental pasting causes limited damage. But based on the accident triangle philosophy, it's best practice to try to cut down on accidental pasting as much as possible, especially when switching contexts or using a collaboratively edited doc such as a Google Doc, or while sharing screen (accidental pasting while within the flow of editing a document locally and privately is not that big a deal).

The main way to cut down on accidental pasting is a mix of (a) improving precision of one's keyboard use in general, and (b) going a little slower whenever using the keys to cut, copy, paste, or bold, e.g., looking at the keyboard rather than touch-typing even if you normally touch-type.

Pause and check before hitting Enter or send

Always pause and check what you're sending before you hit Enter or send, whether you're using email or a messaging tool. In general, I recommend a gap of at least 2-3 seconds between typing your message and sending it, as it gives you enough time to react to anything weird you see in what you're sending.

I generally check the following:

• Is the content of the message what I want it to be?
• Do the links go where they should? (In particular, because links may not be shown as they're replaced by anchor text, this may be important to check).
• Are the recipients the intended ones?

Where applicable, I also try to use preview functionality.

This is an important practice even outside the issue of sensitive data leakage; sending a message to a wrong recipient delays the right recipient receiving it, while also confusing the wrong recipient.

Double-check when sharing documents or putting them in shared folders

Sharing documents in Google Drive or a similar online service for document collaboration is a bit like sending, and the same cautions apply here as they do with sending emails. (Maybe a bit less because you have the ability to edit the document after sending, but you still don't want to share sensitive data).

A silent way of sharing documents is to move them into a shared folder. This generally doesn't proactively notify recipients (but you should double-check this for the specific service you are using) but it still makes the documents accessible to them, and they'll see the documents if they happen to view the folder after you put the document in the folder. Be careful when putting documents in shared folders -- if in doubt, check who has access to the folder, and that it matches the people you want to have access to the document. NOTE: Google does warn you about this when moving a document to a shared folder; in general, it's a good idea to read and ponder such warnings.

Be careful about screensharing

These suggestions apply when screensharing on a conference call (such as Zoom, Google Meet, or Slack Huddle), and also when sharing your screen physically such as in a physical conference room.

• Before screensharing on a conference call, check that you don't have sensitive information in any of the screens that you intend or expect to share. In fact, it may be better to check this when joining the conference call even if you don't need to immediately screenshare, so that you're ready to screenshare whenever.

• If there are services you need to be logged in to for the purposes of the screenshare, it's generally better to log in to them and prepare your relevant screens/tabs prior to the screenshare, so that any information related to your login process (such as your password) don't accidentally leak in the process.

• Check that you don't have sensitive data in your clipboard before sharing screen (in fact, it's good to check before joining the conference call, and to not put sensitive stuff into the clipboard during the call).

• As much as possible, try to share only a specific tab or specific window or specific set of windows. This reduces the risk of sensitive information in other windows getting leaked. Confirm using the software's visual cues that you're sharing the screen(s) you think you're sharing (for instance, Zoom uses a green outline to indicate the windows that you are sharing).

Take additional precautions when using shared computers (such as at a cybercafe or printing shop)

You may sometimes need to use a shared computer, such as one at a cybercafe, for printing and scanning documents. It is possible that software on these machines (such as keyloggers) is surreptitiously collecting information about your login credentials. (I don't know of any incidents where this has happened to me or anybody I know, so it's likely that this is fairly rare).

For more caution, I recommend the following:

• Where possible, prefer to use the option of triggering the relevant actions directly from your own phone or computer, rather than using the cybercafe's computers. For instance, some cybercafes allow you to trigger a print action by emailing a document to a specific address.

• Don't log in to any accounts unless necessary. For instance, if the documents you are printing are public documents, just print them from the public location.

• If the documents to be printed are private, forward them to a non-primary email account that has two-factor authentication but also doesn't have most of your email and isn't tied to your logins on other services, so that even if somehow that account gets hacked, or the information in it gets scraped by the shared computer, your downside is limited. (Even if you didn't do the forwarding beforehand, you may be able to do it using your phone while in the cybercafe). Keep an eye out for notifications about login attempts to that account for the next few weeks.

• After you're done, make sure to log out of the accounts and remove any username or address information.

• Make sure to delete any documents that you downloaded to the shared computer for printing, and also go to the recycle bin to clear it out.

• If you used a web browser, make sure to clear all your activity from the browsing history of the browser.

• As a further best practice, it may be a good idea to change your password on the account you logged in with on the shared machine. Whether to go that far is up to you, and depends on your estimate of the threat level.

Pay attention to warnings and color coding

Many services warn you when they suspect you're sending something suspicious. For instance, Gmail warns you if the email you're trying to send includes links to Google Drive content that the recipients don't have access to. Pay attention to these warnings, as they could help prevent accidental sends to the wrong parties.

Gmail also uses color coding and banners in various ways to indicate cases where it suspects that you're sending a message to the wrong recipients. Pay attention to these colors and banners! For instance, if you're using Google for Workspaces, your work Gmail will color-code the recipients who are outside your organization, so you can eyeball and see that these people aren't in your organization. Similarly, Google Drive issues a warning when you try to share a document with somebody outside your workspace.

In Slack, if you're using Slack Connect to communicate with people from a different workspace, Slack will show a message right above your message-drafting area reminding you that the other person is from a different organization. Pay attention to this.

Figure out how tagging interacts with permissions and notifications

Many services (social media such as Facebook or Twitter, messaging tools such as Facebook Messenger or Slack, collaborative editing tools such as Google Docs) allow for the possibility of tagging people or other entities. Services can differ in the way that tagging interacts with permissions and notifications.

• Some services do not allow for the tagging of people or entities who currently don't have access to the resource. For instance, in Facebook Messenger, only participants in a group message can be tagged; non-participants cannot be tagged.

• Some services allow for the tagging of people or entities who currently don't have access to the resource, and does not notify them. For instance, Slack allows tagging non-participants in group chats or channels, even in cases where the non-participants cannot access even with a link (such as the case of a private channel or group message). Non-participants do not get notified; however, in some cases, Slack may inform the person writing the message that the tagged person or entity is a non-participant and won't get notified (it seems that Slack only does so when it is possible to add the tagged person or entity to the chat or channel).

• Some services notify tagged people or entities that don't have access to the resource. This seems to be the case with Google Docs; if you write a comment on a Google Doc and tag a Google account, that account gets notified even if you don't grant that account access to the document. The account does get a copy of the comment, despite not having access to the document.

In general, it seems safest not to tag entities that you don't want to see the place you're tagging them in, unless you are very sure that the service does not notify them about being tagged.

Reversal strategies

Undo send for emails and messages

• Gmail offers an "Undo send" functionality that is available for a few seconds right after you send an email, if you don't navigate away from the screen. You can configure the send cancellation period for which this "Undo send" is visible by going to Settings > General. Other mail providers may offer similar functionality.

  Note that if you undo send, the message doesn't actually get sent, so recipients don't get it. If you don't undo send, the message get sent after the send cancellation period. Once the message actually gets sent, there is no reversing it.

• Slack allows you to edit or delete messages you've sent any time after you send them; the exact functionality available may depend on your workspace settings. Note that even if you edit or delete the message, the original message would still have been visible to people for the duration from when you sent it to the time you edited or deleted it, and you have no programmatic way to know if they read it or not.

• Facebook's Messenger allows you to "Unsend for everyone" any message (see here); however, the message would have been visible to them prior to unsending. You do get visual cues indicating whether the message was read by them before you unsent it (see here for how to interpret the cues). Even if you unsend, the fact that you sent the message will be included in the message history (in lieu of the message itself).

Undo bad edits to shared documents

When you make a bad edit to a document in Google Docs, Google Slides, or Google Sheets, you should reverse it immediately without navigating away from the tab where you're editing. The reason for this is two-fold:

• Google commits a version of the document every time you switch focus away from it (I don't have an online reference for this; I discovered this by experimentation).

• There is no way to delete old versions from the version history of a document in Google Docs/Slides/Sheets; the only real way is to start afresh with a new document using the "Make A Copy" functionality (but then this document will have a new url and will need to be re-shared with relevant people). For confirmation of this, see here and here.

If nobody else is viewing the document at the time you make your undesired edit, and you reverse the edit immediately without switching tabs, it won't get into the history and will not be visible to others. You should confirm this later by reviewing the history of the document in its most expanded form.

Remove sensitive information from histories

• If you accidentally visited a sensitive url, you should be able to remove it from your browsing history. For instance, see the instructions for Chrome here or for Firefox here. It's best to delete sensitive information as soon as possible so that it doesn't show up in browser autocompletions that might show up later when you're screensharing.

• If you pasted something accidentally into the address bar of your browser and hit Enter, causing you to Google it, you should be able to remove it from your Google search history. See Google's instructions here. It's best to remove from your search history soon so that it doesn't show up as a search suggestion later when you're screensharing with others.

• If you entered sensitive information in your terminal and hit Enter, it may get into your shell history (though it might enter your shell history only after you actually exit that terminal tab). You should be able to delete it from your shell history file (such as .bash_history or .zsh_history). These files are generally only accessible to you, so removing isn't super-urgent but it's probably good to remove so that you don't end up showing it in a screenshare later when searching your history. Similar remarks apply to histories for various REPLs, such as MySQL or Python.

• If you entered sensitive information into your browser console, it gets into your console command history. In Chrome, you can access and edit your console command history by opening devtools of devtools, then manipulating the Local Storage associated with devtools://devtools (either by using Application > Local Storage or using JavaScript in the console using window.localStorage). See here for more related information.

Undo bad edits to posts or comments on online fora

This applies for many online fora; the one I have the most experience with is GitHub.

• Most fora offer the option of fully deleting the post or comment, which removes it from public view. The service may keep a log for a short period.

• Some fora allow you to edit the post or comment and don't show older revisions. In such cases, you can edit out the sensitive information and be done.

• Some fora allow you to edit the post or comment but they do show older revisions. In such cases, you should check whether functionality to delete specific older revisions exists. For instance, GitHub offers the ability to delete specific revisions (except the latest one) for comments on issues and pull requests (see here for more).

It's generally good to investigate what is and isn't allowed by a forum in advance, so that when you do get in the position of having entered sensitive data, you can quickly choose the optimal approach.

Reverse local changes within git repositories

If you made a sensitive change to a git repository that is shared with others (e.g., on GitHub) local changes you make could be reversed if they have not yet been pushed to the remote origin. Here's some guidance:

• If you detect the sensitive data before you commit it locally, you can just undo the change, and then proceed normally. Nothing sensitive got committed, so it won't be in your git history.

• If you detect the sensitive data after you commit it locally but before pushing to the remote origin, you can undo the sensitive data and then amend your commit. Your local git will still have a commit associated with the sensitive data, but it will no longer be associated with the branch. Therefore, it won't get pushed to the remote unless you directly push that specific commit or add it to some other branch that you push. You can later try to remove it through garbage collection or (after you've pushed what you wanted to push) by deleting and refetching the git repository from the remote origin.

• If you already pushed the change remotely, record the commit hash for the commit with sensitive data. Then immediately amend the commit to remove the sensitive data and do a force push (git push --force-with-lease). The commit with sensitive data will still exist on the remote origin, but it's now a dangling commit not associated with any branch (you can confirm this by visiting the page on GitHub and seeing a banner near the top saying so).

  In the case of GitHub, they don't automatically garbage-collect dangling commits, so you need to reach out to their support by clicking "Clear cached views" here. The support can take 1-2 business days to address your request. After that, the dangling commit should not be accessible through GitHub. For more general background on removing sensitive data, see the GitHub doc on the subject. Note that GitHub support does not get automatic access to private repositories when processing a support request; see here for details.

  If any collaborator ran a git fetch or git pull between the time you originally pushed and the time you force-pushed, they'll also have that commit on their local git. However, collaborators who ran git fetch or git pull only after your force-push won't get that dangling commit even if it's still on GitHub.

Detecting and monitoring leakage

What happens if you leak sensitive data and don't even notice that you did so? That's pretty dangerous, because it means that you can't even take the appropriate action to reverse and address it.

Anomaly-checking in the data you publish

I have a few lines of shell script code in one of my scripts (that I run regularly, several times a day) that scan my personal git repositories to make sure they don't include any keywords related to my day job work. This is to address a possibility that I might accidentally type in some work-related stuff into a file in one of my personal repositories. I have some other anomaly checks in a similar spirit.

Things of this kind can be helpful for detecting leaked information that might otherwise be missed.

Activity logs and security alerts for services

If your credential for a service gets leaked, and somebody uses that leaked credential to access the service, it'll be good if you are set up to receive a security alert. For instance, Google logins send a security alert to both the email address that logged in and the recovery email address. Note that this alert will catch when the leak is acted upon by others, and may not catch a latent leak that others haven't acted upon.

Addressing leakage of password or credentials

Change credentials where feasible (after reversing whatever you can reverse)

You should assume that any password or credential is compromised if you think others might have seen it (for instance, you included it in a Slack message and then deleted the message, but aren't sure if people read the message before you deleted it). Therefore, you should change it as soon as feasible. if you are using the same or a highly similar password or credential elsewhere, you should change that too.

This is made easier if:

• You don't use the same or a very similar password/credential in multiple places (as it means you have fewer places to change)

• Your password/credential is not closely tied to personal information about you

• You maintain an organized list of your services/accounts for which you have passwords/credentials, allowing you to audit quickly

Check for any alerts or logs showing unauthorized access over the period where the credentials were leaked

Some services offer details on recent login activity, or the time of recent use of access keys. Take a look at these to check if the leaked credential was used over the time period before it was changed.

Review after being done to make sure you've covered your bases

After you've done whatever reversal and credential changes you need to, you should sit down and review the situation, systematically going over what happened to see if things are back to a secure state. If there are other trusted people with whom you can discuss the situation, please do so.

Addressing leakage of factual information

Check if others accessed the information

This is most relevant to factual information that, unlike credentials, cannot simply be changed.

In case of factual information that others were not supposed to have, you may be able to check if they were able to access the information without explicitly asking them. Some cases:

• If you're using Google through a work or school account, then for documents accidentally shared with others in Google Drive, you can go to Tools > Activity dashboard > Viewers to see who all has viewed the document. After you've fixed the permissions, you should be able to check this. If the people you accidentally shared with did not access the document, you can be confident that it didn't leak to them. This functionality is not available for personal Google accounts; see here for confirmation.

• Facebook Messenger has visual cues to show if somebody read a message (see here for how to interpret the cues). If you delete the message prior to them having read it, you can be reasonably confident they didn't access it.

Figure out what kind of secret data it was from the person it leaked to

The benign case of not-very-actionable information they don't and shouldn't know but they know about

An example here is that you accidentally share information about your company's quarterly revenue with a work colleague who is not supposed to have access to the data. Then you detect and delete the data. There's nothing particularly surprising about the data you accidentally shared; it's just that as a matter of policy, your colleague shouldn't have access to that data. Your colleague is also aware that such data exists, and that you have access to the data -- the only thing they didn't know is the specifics.

The trickier case of information that is genuinely surprising and actionable to them

Let's say you are a senior executive at a company and have been tasked with figuring out what half of the people in your division to fire. If you accidentally share musings/thoughts related to this with one of the people under you (who might end up being in the line of fire, even if it's unclear right now), this is a case of information that is actionable to them. This is a tricky situation to be in, and here you have to exercise a judgment call regarding whether to take them into confidence explicitly versus go with the risk that they might have seen the information.

In this case, the other person may feel stress and fear about the situation. Depending on how they model you, and what they expect from you, they may also feel disappointed in you for not sharing the information with them earlier. Finally, if you only leaked to a subset of the affected people, you may create asymmetries between them (e.g., now some of the people who may be fired know they may be fired, and others don't) and the possibility of further leaks.

Explicitly discuss if you're sure they saw it

If you're sure the other person saw it, it seems best to explicitly discuss.

In the case of "not-very-actionable information they don't and shouldn't know but they know about" it may be as simple as saying "Hey, I accidentally leaked this info to you; you may have seen it; please ignore and delete it. Sorry!"

In the "trickier case of information that is genuinely surprising and actionable to them", you'll need more of a strategy. The exact strategy you choose partly depends on the reasons you kept the information secret in the first place.

• In some cases, it makes sense to talk to the people who've seen the information and ask them to keep it in confidence for now.

• In other cases, it may make sense to openly share with a wider pool (if you're concerned about asymmetries between the subset of people who know and the others who don't know). This can tamp down on speculation and the possibility of ongoing leaks.

Make a judgment call about explicit sharing if you aren't sure if they saw it or not

If you were able to re-hide factual information that you temporarily leaked, and aren't sure if people saw it during the interim, you have to make a judgment call: do you operate under the assumption that they probably didn't see it (and it's as good as if you had never leaked it), or do you act on the assumption that they might have seen it? The specifics here will vary based on the situation.

In the "benign case of not-very-actionable information they don't and shouldn't know but they know about" it doesn't matter too much; it may be easiest to just not say anything, but if they bring it up you can tell them you accidentally shared it and ask them to ignore or delete it.

In the "trickier case of information that is genuinely surprising and actionable to them", you have to choose between sharing with them and asking them to keep confidence, sharing more widely with all affected parties, and just keeping quiet until you have stronger evidence that they know.

What can you do in advance to make addressing leakage easier?

When hiding sensitive information over an extended period of time, you should assume some probability of leakage. Here are a few thoughts on how you can make it easier:

• Don't make enemies! In general, people are more likely to exacerbate a leakage (by leaking it further) if they don't like you. Don't give people reasons to hate you or want to get back at you.

• As much as possible, try to minimize the number of cases where you're hiding from people "information that is genuinely surprising and actionable to them"; as much as possible, data that you hide should be of the form of "not-very-actionable information they don't and shouldn't know but they know about" (for instance, detailed financial statements that are consistent with the high-level picture they have of finances, but are being kept secret). As long as most secrets aren't things that are very actionable to the people who they get leaked to, these people will just ignore them (unless they hate you, which brings me back to the "Don't make enemies!" point).

• If a secret needs to be maintained for an extended period of time, invest in some upfront organization around how to make sure that stuff related to that secret doesn't leak. For instance, maybe only do things related to that secret during specific times of the day when you are not interacting with other people.

• Formulate contingency plans for how to deal with a leakage. In particular, think through what path you'll take if the information has definitely or probably leaked, per the preceding subsections. Think whether the fact that you had the information and kept this a secret will seem, in hindsight, to be a moral failing on your part. If it will, examine whether it really is the right thing to keep it secret.

More on the point about not making enemies, and the need for extra caution if making enemies is inevitable

I want to dig in on the point about not making enemies because it's pretty important when thinking about the impact of any kind of mistake, whether it's data leakage or anything else. When you have enemies, there's more risk that they'll pounce on and exploit your mistakes. In contrast, if people are friendly with you and like and respect what you're doing, they're more likely to be generous to your faults, particularly if those faults don't directly hurt them.

There are, however, cases where the very nature of what you're doing means the other side is going to be hostile to you, and you have accepted that. In that case, you need to invest in a substantially greater level of caution as the cost of data leakage is higher. One example of such a situation is when you're a whistleblower for a terrible situation. It's very likely that the fact that you're blowing the whistle itself needs to be kept a secret from the people whose whistle you're blowing.

Or, there could be cases where people don't have enmity against you personally, but you're a messenger of really bad news coming from elsewhere (for instance, you have information in advance about mass layoffs and have been told to not communicate this to any of the people in the pool that might be considered for layoffs). Here again, you can mitigate the impact to some extent by not additionally behaving in ways that make you personally unlikable. But given the base animosity of the situation, exercising additional caution is probably very important.

Meta comments

Why think about this in advance? Object-level and meta-level reasons

Why think about this sort of thing in advance, rather than wait to let a problem happen and then tackle it after it happens?

I think there are several object-level reasons to think about this in advance:

• Many of the best practices are things that need to be done in advance: For instance, setting the right send cancellation period so that you can undo send is not something you can do after sending an email accidentally.

• Speed matters a lot in addressing data leakage: The time it takes to investigate things, even if a few minutes, can make all the difference. For instance, figuring out how to delete sensitive data from git commits is somewhat nontrivial and can take several minutes to figure out.

• Dealing with a data leakage is stressful and you may miss doing "obvious" things if you don't have prior preparation: For instance, ideally, if you leak a credential, you should both delete the leakage and change the credential. But it may not occur to you in the heat of the moment -- you may just end up deleting the leakage but not changing the credential. Or, if you accidentally leak a credential in a GitHub comment, you may just edit the comment to remove the leaked value, and forget that GitHub has an edit history and the old version needs to be explicitly deleted.

There are also meta-level reasons that might be relevant for the LessWrong audience; data leakage is a concrete example of a low-frequency but high-cost disaster, and as such, we don't get a lot of day-to-day feedback around it. Thinking about this sort of thing offers relatively easy practice of security mindset as well as practice in reducing accidents. Such practice could be useful for avoiding even rarer and higher-stakes problems.

Tool-specific guidance versus the general ideas

In this post, I've included concrete details for several specific tools such as Gmail, Messenger, Slack, git and GitHub, Google Docs. These tools may change over time and some of my advice may become outdated. Also, you might be using very different tools, so you may not be able to put the guidance to immediate use.

I wanted to put in specific guidance for tools that people are likely to use in order to make the action items fairly concrete and make the post useful. But I also think that they are useful even if you don't happen to use these specific tools, but they provide a framework for thinking about the structure and design choices underlying these tools. Even if you use a different tool, it was likely designed with relatively similar design constraints.

For instance, Gmail, Facebook Messenger, and Slack all offer different versions of "undo send" that are all meaningfully different in their implications for how to deal with sensitive data:

• Gmail is best in that if you can undo send within the send cancellation period, nobody else could have seen it. But it's also worst in terms of having a really small send cancellation period, after which the cat is out of the bag, you can't undo send, and you can't even know if others have seen it.

• Messenger is best in that you can undo send at any time and you can know if the other person saw the message as of when you undo send. But it's worst in terms of permanently showing that you unsent the message, and it's also bad in that there is no "no-regret" cancellation window (unlike Gmail's short send cancellation period).

• Slack is best in that you can completely delete the message leaving no trace of it, and you can do so at any time. But it's worse than Messenger in terms of not giving feedback on whether others saw your message, and it's worse than Gmail in terms of having no "no-regret" cancellation window.

If you happen to use something different from all three of these, you at least have a framework and a set of comparison points, which can lead you to ask the right questions about what your service supports and what implications it has for the leakage of sensitive data.

Am I encouraging people to hide factual information more efficiently?

I think the idea of hiding credentials is relatively uncontroversial; this is good for security (even if others have access to the same resources you do, using separate credentials is better for security as it allows you to swap your credentials out without affecting others).

On the other hand, hiding factual information from others is not always good. It's clearly necessary in at least some cases, but we could also argue that there is a wide range of cases where it's done in service of wrong purposes. One could even argue that in some cases, it's better for the world if the information being hidden did get leaked.

I think that even though the post offers guidance on how to hide factual information more effectively, it also raises considerations that should encourage people to reduce reliance on hiding relevant factual information for wrong purposes. In particular, earlier in the post, I wrote:

• Don't make enemies! In general, people are more likely to exacerbate a leakage (by leaking it further) if they don't like you. Don't give people reasons to hate you or want to get back at you.

• As much as possible, try to minimize the number of cases where you're hiding from people "information that is genuinely surprising and actionable to them"; as much as possible, data that you hide should be of the form of "not-very-actionable information they don't and shouldn't know but they know about" (for instance, detailed financial statements that are consistent with the high-level picture they have of finances, but are being kept secret). As long as most secrets aren't things that are very actionable to the people who they get leaked to, these people will just ignore them (unless they hate you, which brings me back to the "Don't make enemies!" point).

[...]

• Formulate contingency plans for how to deal with a leakage. In particular, think through what path you'll take if the information has definitely or probably leaked, per the preceding subsections. Think whether the fact that you had the information and kept this a secret will seem, in hindsight, to be a moral failing on your part. If it will, examine whether it really is the right thing to keep it secret.

I think these points probably push in the direction of not keeping material, actionable secrets from people for the wrong reasons.