Keeping secrets out of logs (2024)
allan.reyes.sh
allan.reyes.sh
Just an excellent example of how to approach & elucidate a problem domain.
At first sight it seems a complicated and inferior approximation of techniques from the article: not automatically single use, not statically checked, somewhat error prone for proper secret usage, not really preventing well-intentioned idiots from accidentally extracting, "laundering" and leaking the secret, removing secrets from logs at a dangerously late stage with some chance of leaks.
But one big reason I suggested this technique is that you want the object to keep protection on the String while having it look and feel as much like the underlying contents as possible, so that the final unsealing can occur as little (& as late) as possible. The more warts you put around your secret, the less usable it will be. You thought you made the Secret "single-use", but what you really did was to just encourage someone to keep the unsealed String around and reuse that, because you gave them a Secret type and they needed a String type. And now you have no way to detect if they accidentally log it, or throw an exception with some local variable containing it. Whereas this technique would still immediately catch any leakage in those cases.
Again: this technique is a supplement, not a substitute. You absolutely should still add static checks where you can. Have your Secret type too. The point here is that your Secret.unseal() method can still return a String that is useful for callers while offering you some protection on the value, instead of instantly going from protected->unprotected and exposing the contents with zero protection.
> You thought you made the Secret "single-use", but what you really did was to just encourage someone to keep the unsealed String around and reuse that, because you gave them a Secret type and they needed a String type
And an exact match is just part of the problem; if a dev redacts the end and another dev redacts the start, you can still reassemble the secret with enough logs.
Regex matching on logs is slow but if performed on every node the CPU load is distributed vs. doing this upstream. Configuration management can push the regex rules to all the nodes. This won't help with unknown-unknowns but those can be added quickly to all nodes through configuration management after peer review.
Rsyslog also supports encrypting the log stream so that secret leakage is limited to the sending nodes and the central nodes and it checks a few boxes.
Another thing that helps is limiting to warn and above sent upstream and using an agent on the local nodes to monitor for keywords in the range of info to debug to let someone know to go check the node logs. Less junk on the centralized servers that may have SOC1/SOC2/PCI/FEDRAMP log retention requirements. One can not leak what is not sent in the first place.
Also if you have audit records, you want accessing a secret to be logged separately from accessing logs.
You could have 100s of people who have a business need to look at syslog from a router, but approximately nobody who should have access to login creds of administrative users and maybe 10s of people with access to automation role account creds.
The kitchen sink example in particular is one that trips up people. Without knowing the specifics of how a library may deal with failure edge cases, it can catch you off guard (e.g., axios errors including API key headers).
A lot of these problems come from architectures where secrets go over the wire instead of just using signatures/ids. But in cases where you have to use some third party platform, there's often no choice.
What I did at a previous shop was remove the passwords as part of a smart gdb script that runs when the core is dumped, before it gets written to a readable location.
Writing the script also helped to demonstrate how to extract the passwords in the first place.
> And while people will write the code that accidentally introduces sensitive data into logs, they’re also the ones that will report, respond, and fix them.
This should probably be the first point and not the last.
> const secret = new Secret("...")
one of those things that's obvious in retrospect. That's a cute trick I'll definitely be stealing.
Which reminds me of why I hate tiny standard libraries as seen in JavaScript: features like SecureString work only if they're used pervasively. It has to be in the std lib and it has to be used everywhere so that you almost never have to unwrap them. It's critical that credentials are converted to SecureString as soon as possible and that they stay as SecureString values until the last possible instant when they're passed to some external API call deep inside even a third-party a library.
We’ve been working on an open source tool, Kingfisher, that pairs fast detection (Hyperscan + Tree-Sitter) with live validation for a bunch of providers (cloud + common SaaS) so you can down-rank false positives and focus on the secrets that really matter. It plugs in at the chokepoints this post suggests: CI, repo/org sweeps, and sampled log archives (stdin/S3) after a Vector/rsyslog hop.
Examples:
kingfisher scan /path/to/app.log --only-valid
kingfisher scan --s3-bucket my-logs --s3-prefix prod/2025/09/
Repo: https://github.com/mongodb/kingfisher (Apache-2.0)
Disclosure: I help maintain Kingfisher.
I have very strong opinions on this issue that boils down to. _why are you logging everything you lazy asses_ and _adding all the secrets into another tool just to scan for them in logs just adds another point for them to leak_...
Especially since the ability of lines getting censored even when the secrets were just part of words showed that probably no hashing was involved.
But its a security tool so it stays. I kinda feel like Cassandra but I think I can already predict a major security issue with it or others with the same functionality in the future. its like some goddamn blind spot that software that is to prevent X cannot be vulnerable to X but somehow often is vulnerable because prevention of X and not being vulnerable to X are two separate things somehow.
Logging "everything" could include stack traces and parameter values at every function call. Take the information you can get from a debugger and imagine you log all of it. Would that be necessary to determine why a defect is triggered?
Second, "lazy":
Logging has many useful aspects, but it is also only a step or two above adding print statements to the code, which again leads to the "lazy." If you have the inputs, you should be able to reproduce the execution. The exceptions include "poorly" modularized code, side effects, etc.
Alternatives.
I've found it helpful for complex failures to make sure that I include information about the system. For example, the program couldn't allocate memory: Was it continuous chunks of memory or a memory leak? How much free memory is there, versus the shapes of the free memory (Linux memory slabs)? What can I do to reset this state? (reboot was the only option)
Finally, a quote a colleague shared with me when I once expressed my love of logging. In the context of testing online games:
"Developers seem drawn to Event Recorders like moths to a flame. Recording all game/ mouse/ network/ whatever events while playing the game and playing them back is a bad idea. The problem is that you have an entire team modifying your game's logic and the meaning or structure of internal events on a day-to-day basis. For The Sims Online and other projects, we found that you could only reliably replay an event recording on the same build on which it was recorded. However, the keystone requirement for a testing system is regression: the ability to run the same test across differing builds. Internal Event Recorders just don't cut it as a general-purpose testing system. UI Event Recorders share a similar problem: when the GUI of the game shifts, the recording instantly becomes invalid."
Page 181, "Section 2.1 Automated Testing for Online Games by Larry Mellon of Electronic Arts", in Massively multiplayer game development 2, edited by Thor Alexander, 2005
I've never used axiom, but all the logging platforms I've used like splunk, datadog, loggly are a major op-ex line item.
And telling your developers their time is priceless means they will produce the lowest quality product.
Secondly, you can't represent the heap & stack well as strings. Concurrent threads and object trees are better debugged with a debugger (e.g. gdb).
Even if I trust me.
Audits happen. I assume other people will eventually see this bad practice.
My argument is that generally everyone has access to all the logs. If you restrict the access and add guardrails around it, you can minimize the surface area and also ways it can be leaked out.
If you take a defensive approach towards, you have to assume that some secret is getting logged somewhere. The goal then becomes a way to reduce the surface area or blast radius of this possible leakage.
Some of the same techniques apply, like using domain primitives, but some PII (like names and addresses) is eventually templated into flatter (text) values, and processed by other layers which do not recognize 'brands' as suggested.
Data scanners: Regexes are fine for SSNs and the like, but to be really effective, one would need a full-on Named Entity Recognition in the pipeline, perhaps just as a canary. (Wait, that might actually work?)
Dataflow analysis and control applies in a BIG way, e.g. separating an audit log for forensics, where you really NEED the PII, from a technical log which the SREs can dig into without being suspected of stealing sensitive info. Start there.
secrets.forEach(secret => logMessage = logMessage.replaceAll(secret, '**'))
secrets can also churn, so even if you did your example would require something besides an in-memory array.
and, the final point: what if your secret masking code fails on an exception, too ;)