You built it because you wanted to share passwords:
And your flow is: I encrypt my password; I upload the encrypted password to your server.
And I share the password to the encrypted password as plain text.
Why do I have to upload the encrypted password to your server, and not just use signal disapearing messages, or telegram secure channel disappearing messages to share the encrypted password there.
And I can use any other side channel to share the second password, like whatsapp, or regular plain mail.
It feels to me that you made a two step process into a one step process but increased the risk by adding you in the middle.
Why would I offload my trust to you instead of doing the second step?
Your skepticism is valid and if your flow already includes: A secure messaging tool (e.g. Signal), a GPG workflow or local encryption or a team that uses shared password vaults. Then to be fair Stasher might not be better.
I built Stasher for me. I wanted an easy, CLI-first way to share one-time secrets without worrying about accounts, apps, or trust. If Signal or GPG works better for you that’s totally cool.
Stasher exists to make casual, secure sharing simpler not to replace tools you already trust.
I’d also add the language to the mix. I know you can write good code with TS/JS, but the dependency surface is just so large, I’m not comfortable with security code written in it yet (maybe at some point). Add that the repositories were created in the past week, so we can’t see the actual dev practice (was it all vibe coded? What bugs were there?).
I hadn’t considered your second point, but even the authors GH account has an AI picture. I have no idea who this person is or what online/HN reputation they have.
Thanks for raising these concerns — totally fair in the context of security tools.
I’m not anonymous, just cautious. I’m a solo builder, and this is a focused identity for the project. In fact, that's why I implemented full supply chain transparency from day one: signed releases, SLSA attestations, SBOMs, and Rekor logs. You don't need to trust me you can see the code for your self.
Ultimately, you're right — if you can't verify it, you shouldn't trust it.
That’s the whole point of the system: zero trust and verifiable cryptographic guarantees.
A "focused identity" with no links to other identities is anonymous by definition.
More importantly, this project is not "zero trust" and calling it such is borderline deceptive.
I can verify the artifacts you're shipping contain the code in the repo (or I could just clone the repo myself), but I cannot automatically verify that your code is non-malicious and free of bugs. That is what I am trusting when using your software, and I have serious doubts about the "free of bugs" part for AI generated software.
I’m right there with you in mistrusting AI generated code but - you also can’t automatically verify that human-written code is non-malicious and bug free.
Cryptography/security is a trust business. Without some kind of personal (or even project) history, I know nothing about you or the project. And if I can’t verify you, I can’t trust you. The rest doesn’t matter much to me.
It means the releases are cryptographically signed using GitHub OIDC, with SLSA v1 provenance and entries in the Rekor transparency log.
That means:You can verify every artifact against its source code i.e I have not tampered with the code post deployment. for example part of the build is a dry-run on the worker build, this is stored as part of the build so you can see / confirm the exact code that was uploaded and this code is signed.
What people mean with "trust" here is whether they trust there are no subtle security issues. While I think "don't roll your crypto" has been somewhat overstated at times (someone has to write crypto), there is certainly some truth in that you need to be very careful writing this code and that mistakes are incredibly easy to make, even for competent developers.
If I were to release something like this then people can see that 1) this is a guy with >20 years of various contributions to open source and he seems like a basically competent guy we can trust (as much as you can ever trust a single person), but also that 2) he's not a crypto guy and there may very well be oversights. Maybe there are none, but you know...
If someone like, say, Filippo Valsorda would release someone like this, then people could see that 1) is basically the same as me, and 2) he's also a well known crypto guy with a good track record. This is not a guarantee there are no oversights, but I would certainly be surprised if there were major ones. I would certainly trust that more than anything I would write.
The whole signing stuff is kind of a red herring IMO. I mean, it's not bad to have I guess, but honestly I don't really care. If anything, focusing to strongly on "box ticking security" so early on seems like the wrong thing to focus on.
We'll find out in an hour, but I bet openai trained em-dashes out of gpt-5 and that will confuse a lot of people. At least you'll be able to write the way you want to...
Humans also use em dashes — like that. My browser for one automatically creates them on HN if you correctly type a space, two hyphens then another space. Maybe the dude just has good grammar.
Hey. Only the ciphertext is stored on the server; the key never leaves your machine. The uuid:key format is just a pointer to the encrypted payload. Without the key, the server’s stash is useless. Zero-knowledge by design
Reactive expiry — When someone tries to retrieve a stash (destash), the Durable Object checks the creation timestamp before serving. If it's older than 10 minutes, it refuses the request.
Proactive cleanup — Every stash’s Durable Object sets a scheduled alarm to self-destruct after 10 minutes. This removes the coordinating DO and ensures the encrypted blob in KV expires (via TTL).
So even if someone tries to cheat the system, or access after the 10-minute window, they’ll get an error — the stash is gone.
This is part of what makes it “burn-after-read, or expire-after-time”. No guessing, no timers in memory or cron job workers.
How are salts handled?
Stasher uses AES-256-GCM, which does not require a traditional salt like in password hashing (e.g. PBKDF2, bcrypt). Instead, it uses an IV (initialization vector).
With a fresh 96-bit IV is generated for every encryption
AES-GCM uses that IV as part of the encryption process, ensuring non-deterministic ciphertext. The IV is not secret, and is uploaded alongside the ciphertext and GCM tag
On decryption, the IV is used to reconstruct the exact same cipher context
So in short: No static salts and no reused IVs
Everything you need to decrypt is bundled with the encrypted stash, except the key, which stays with the user (as part of the uuid:base64key token)
Yes, that's a fair comment technically speaking: Cloudflare Workers + KV + Durable Objects is a backend. I was trying to imply No user accounts, no persistent database, no stateful sessions etc I will reword - thanks for the feedback
I wish it was easier to run code in browser that you could know did not make any network connections, thinking mostly of the client creating secrets here.
That requires a dedicated instance of your browser as (AFAIK) most browsers don't support per-tab proxy configuration. If I understood correctly, parent wants tabs to work normally but offline tabs (like the secret generator) to be airgapped.
Yes, you are understanding it correctly, the server (Cloudflare Worker + Durable Object + KV) in Stasher is only needed to enforce the burn-after-read behaviour
zero setup, burn after read, no key exchange required, GPG is ideal for persistent trust relationships (e.g., signing emails), Stasher is purpose-built for temporary relationships. To me GPG is overkill for sharing simple shares. Defo not trying to replace GPG, just a different use case.
>Author's English (when not written by a LLM) sounds translated
Doesn't inspire confidence.
Use GPG, it's not difficult. For non-technical folks, use signal or disappearing messages. For slightly more secure comms with non-technical people, use a combination of rot13 / caesar / similar.
Heh, maybe not for you and me, but if you've never tried to coach a less-technical person through setting up GPG and then using to encrypt/decrypt, it's an eye-opening experience
You built it because you wanted to share passwords:
And your flow is: I encrypt my password; I upload the encrypted password to your server.
And I share the password to the encrypted password as plain text.
Why do I have to upload the encrypted password to your server, and not just use signal disapearing messages, or telegram secure channel disappearing messages to share the encrypted password there.
And I can use any other side channel to share the second password, like whatsapp, or regular plain mail.
It feels to me that you made a two step process into a one step process but increased the risk by adding you in the middle.
Why would I offload my trust to you instead of doing the second step?
Your skepticism is valid and if your flow already includes: A secure messaging tool (e.g. Signal), a GPG workflow or local encryption or a team that uses shared password vaults. Then to be fair Stasher might not be better.
I built Stasher for me. I wanted an easy, CLI-first way to share one-time secrets without worrying about accounts, apps, or trust. If Signal or GPG works better for you that’s totally cool.
Stasher exists to make casual, secure sharing simpler not to replace tools you already trust.
Yes, valid, congratulations on shipping!
It's just that the entry level for adopting a new tool (for other people) is:
Convince my recipient to use this system instead of "Why not just send the password as we usually do on our secret chat."
And then we spend 20 minutes talking about it and me advocating for their unknown and unaccountable creator.
The best tool I've found for password dissemination (you can create with it too, or not) is Password Pusher
https://github.com/pglombardo/PasswordPusher
East to deploy, and just works, nice set of sane features and defaults.
I'm sorry, but I would never use this because of two major dealbreakers (and I would encourage others to exercise serious caution as well):
1. Code is largely if not entirely written by AI
2. Author is completely anonymous, using a dedicated GitHub and HN account for this specific project
Both of these are really bad for security-sensitive software.
I’d also add the language to the mix. I know you can write good code with TS/JS, but the dependency surface is just so large, I’m not comfortable with security code written in it yet (maybe at some point). Add that the repositories were created in the past week, so we can’t see the actual dev practice (was it all vibe coded? What bugs were there?).
I hadn’t considered your second point, but even the authors GH account has an AI picture. I have no idea who this person is or what online/HN reputation they have.
Thanks for raising these concerns — totally fair in the context of security tools.
I’m not anonymous, just cautious. I’m a solo builder, and this is a focused identity for the project. In fact, that's why I implemented full supply chain transparency from day one: signed releases, SLSA attestations, SBOMs, and Rekor logs. You don't need to trust me you can see the code for your self.
Ultimately, you're right — if you can't verify it, you shouldn't trust it.
That’s the whole point of the system: zero trust and verifiable cryptographic guarantees.
Appreciate the scrutiny
A "focused identity" with no links to other identities is anonymous by definition.
More importantly, this project is not "zero trust" and calling it such is borderline deceptive.
I can verify the artifacts you're shipping contain the code in the repo (or I could just clone the repo myself), but I cannot automatically verify that your code is non-malicious and free of bugs. That is what I am trusting when using your software, and I have serious doubts about the "free of bugs" part for AI generated software.
I’m right there with you in mistrusting AI generated code but - you also can’t automatically verify that human-written code is non-malicious and bug free.
Cryptography/security is a trust business. Without some kind of personal (or even project) history, I know nothing about you or the project. And if I can’t verify you, I can’t trust you. The rest doesn’t matter much to me.
But maybe that’s just me.
I get it. An 'anonymous' author is a deal breaker for some. I respect that.
The repo is public. The releases are signed. The attestations are published. Nothing hidden.
If that’s not enough — totally fair and I am sure many others would agree. Appreciate your point of view and taking time to give feedback.
Would you please elaborate on how "the releases are signed" helps establish trust in the context of your anonymous developer account?
It means the releases are cryptographically signed using GitHub OIDC, with SLSA v1 provenance and entries in the Rekor transparency log.
That means:You can verify every artifact against its source code i.e I have not tampered with the code post deployment. for example part of the build is a dry-run on the worker build, this is stored as part of the build so you can see / confirm the exact code that was uploaded and this code is signed.
What people mean with "trust" here is whether they trust there are no subtle security issues. While I think "don't roll your crypto" has been somewhat overstated at times (someone has to write crypto), there is certainly some truth in that you need to be very careful writing this code and that mistakes are incredibly easy to make, even for competent developers.
If I were to release something like this then people can see that 1) this is a guy with >20 years of various contributions to open source and he seems like a basically competent guy we can trust (as much as you can ever trust a single person), but also that 2) he's not a crypto guy and there may very well be oversights. Maybe there are none, but you know...
If someone like, say, Filippo Valsorda would release someone like this, then people could see that 1) is basically the same as me, and 2) he's also a well known crypto guy with a good track record. This is not a guarantee there are no oversights, but I would certainly be surprised if there were major ones. I would certainly trust that more than anything I would write.
The whole signing stuff is kind of a red herring IMO. I mean, it's not bad to have I guess, but honestly I don't really care. If anything, focusing to strongly on "box ticking security" so early on seems like the wrong thing to focus on.
Thanks for your feedback 'I tried to use of the shelf stuff for the crypto and utilised what I believe to be battle tested.
CLI uses node.js built-in crypto module only -randomBytes - createDecipheriv - createCipheriv.
Web app uses Web Crypto API only.
I'll send Filippo a Postcard and see if he will review it :-P
Is this a bit?
I also now see that you're using em dashes in your replies - are even the HN comments AI generated???
Everyone please just stop with the em dash hysteria. You just tried to use one yourself — apparently you just don’t know how to type it.
We'll find out in an hour, but I bet openai trained em-dashes out of gpt-5 and that will confuse a lot of people. At least you'll be able to write the way you want to...
I use emdashes like this-- see the difference
Humans also use em dashes — like that. My browser for one automatically creates them on HN if you correctly type a space, two hyphens then another space. Maybe the dude just has good grammar.
Some colleagues use LLMs to translate their messages to English. Same can ve applied here
I feel like I’d rather send “uuid:cipertext” so the cipertext never touches a server, but logically the security seems the same.
Hey. Only the ciphertext is stored on the server; the key never leaves your machine. The uuid:key format is just a pointer to the encrypted payload. Without the key, the server’s stash is useless. Zero-knowledge by design
I feel like I'm being very stupid. If the key never leaves my machine, how do I share a secret?
When you run:
npx enstash "my secret"
Stasher performs everything locally:
Generates a random 256-bit encryption key
Encrypts your secret using AES-256-GCM
Sends only:
the ciphertext
the IV (initialization vector)
the auth tag
a randomly generated UUID
The encryption key is never sent to the server. It never leaves your machine.
You are then shown a single string:
uuid:base64key
The uuid points to the encrypted stash on the server
The base64key is the encryption key you just generated
Only the person who has both parts can decrypt the secret
How You Share the Secret
You send the full uuid:base64key token to your recipient — over any channel you like slack or whatever.
When they run:
npx destash "uuid:base64key" on the token
Stasher:
Fetches the encrypted stash using the uuid
Deletes it immediately (burn-after-read)
Decrypts it locally using the base64key
Shows the secret
The server never sees the key. Not during upload or during retrieval.
>no accounts, no logins, no servers to trust.
>The uuid points to the encrypted stash on the server
No servers… “on the server.” hmmm, I must be missing something.
So the ten minute thing is a trust issue. How are salts handled?
So the ten minute thing is a trust issue?
Stasher enforces expiry in two layers:
Reactive expiry — When someone tries to retrieve a stash (destash), the Durable Object checks the creation timestamp before serving. If it's older than 10 minutes, it refuses the request.
Proactive cleanup — Every stash’s Durable Object sets a scheduled alarm to self-destruct after 10 minutes. This removes the coordinating DO and ensures the encrypted blob in KV expires (via TTL).
So even if someone tries to cheat the system, or access after the 10-minute window, they’ll get an error — the stash is gone.
This is part of what makes it “burn-after-read, or expire-after-time”. No guessing, no timers in memory or cron job workers.
How are salts handled?
Stasher uses AES-256-GCM, which does not require a traditional salt like in password hashing (e.g. PBKDF2, bcrypt). Instead, it uses an IV (initialization vector).
With a fresh 96-bit IV is generated for every encryption
AES-GCM uses that IV as part of the encryption process, ensuring non-deterministic ciphertext. The IV is not secret, and is uploaded alongside the ciphertext and GCM tag
On decryption, the IV is used to reconstruct the exact same cipher context
So in short: No static salts and no reused IVs
Everything you need to decrypt is bundled with the encrypted stash, except the key, which stays with the user (as part of the uuid:base64key token)
I'd recommend changing your tagline -
> Share secrets from your terminal. One-time only. No accounts. No backend. No BS.
A server sure sounds like a backend to me.
Yes, that's a fair comment technically speaking: Cloudflare Workers + KV + Durable Objects is a backend. I was trying to imply No user accounts, no persistent database, no stateful sessions etc I will reword - thanks for the feedback
I wish it was easier to run code in browser that you could know did not make any network connections, thinking mostly of the client creating secrets here.
whats preventing setting up a proxy (like mitmproxy, burpsuite interceptor) in the browser? pretty easy
That requires a dedicated instance of your browser as (AFAIK) most browsers don't support per-tab proxy configuration. If I understood correctly, parent wants tabs to work normally but offline tabs (like the secret generator) to be airgapped.
Do I understand this correctly that the server here is only needed to make sure the secret it's only read once?
Yes, you are understanding it correctly, the server (Cloudflare Worker + Durable Object + KV) in Stasher is only needed to enforce the burn-after-read behaviour
What does "burn-after-read" mean? Just that it can't be retrieved a second time?
Exactly. Once it's read it's gone. If it's not read within 10 minutes... It's gone.
Why use this instead of GPG?
zero setup, burn after read, no key exchange required, GPG is ideal for persistent trust relationships (e.g., signing emails), Stasher is purpose-built for temporary relationships. To me GPG is overkill for sharing simple shares. Defo not trying to replace GPG, just a different use case.
>LLM generated
>Buzzwords
>Author's English (when not written by a LLM) sounds translated
Doesn't inspire confidence.
Use GPG, it's not difficult. For non-technical folks, use signal or disappearing messages. For slightly more secure comms with non-technical people, use a combination of rot13 / caesar / similar.
> Use GPG, it's not difficult.
Heh, maybe not for you and me, but if you've never tried to coach a less-technical person through setting up GPG and then using to encrypt/decrypt, it's an eye-opening experience
wouldnt that command line output be sent to .bash_history of the logged in user?
you probably want to unset HISTFILE and then set +o history before sending these commands to bash
Great point — I’m planning to add a --stdin option explicitly for cases like this. Thanks for raising it. I will add to the readme in the meanwhile.