I have yet to see such a company that's so insecure that they would keep their CLI closed source even when the secret sauce is in the model that they control already and is closed source.
Not only that, wouldn't allow other CLIs to be used either.
There are now several comments that (incorrectly?) interpret the undercover mode as only hiding internal information. Excerpts from the actual prompt[0]:
NEVER include in commit messages or PR descriptions:
- The phrase "Claude Code" or any mention that you are an AI
- Co-Authored-By lines or any other attribution
BAD (never write these):
- 1-shotted by claude-opus-4-6
- Generated with Claude Code
- Co-Authored-By: Claude Opus 4.6 <…>
This very much sounds like it does what it says on the tin, i.e. stays undercover and pretends to be a human. It's especially worrying that the prompt is explicitly written for contributions to public repositories.
I would have expected people (maybe a small minority, but that includes myself) to have already instructed Claude to do this. It’s a trivial instruction to add to your CLAUDE.md file.
It's less about pretending to be a human and more about not inviting scrutiny and ridicule toward Claude if the code quality is bad. They want the real human to appear to be responsible for accepting Claud's poor output.
That’s how I’d want it to be honestly. LLMs are tools and I’d hope we’re going to keep the people using them responsible. Just like any other tools we use.
None of this is really worrying, this is a pattern implemented in a similar way by every single developer using AI to write commit messages after noticing how exceptionally noisy they are to self-attribute things. Anthropics views on AI safety and alignment with human interests dont suddenly get thrown out with the bathwater because of leaked internal tooling of which is functionally identical to a basic prompt in a mere interface (and not a model). I dont really buy all the forced "skepticism" on this thread tbh.
You can already turn off "Co-Authored-By" via Claude Code config. This is what their docs show:
~/.claude/settings.json
{
"attribution": {
"commit": "",
"pr": ""
},
The rest of the prompt is pretty clear that it's talking about internal use.
Claude Code users aren't the ones worried about leaking "internal model codenames" nor "unreleased model opus-4-8" nor Slack channel names. Though, nobody would want that crap in their generated docs/code anyways.
Seems like a nothingburger, and everyone seems to be fantasizing about "undercover mode" rather than engaging with the details.
The name "Undercover mode" and the line `The phrase "Claude Code" or any mention that you are an AI` sound spooky, but after reading the source my first knee-jerk reaction wouldn't be "this is for pretending to be human" given that the file is largely about hiding Anthropic internal information such as code names. I encourage looking at the source itself in order to draw your conclusions, it's very short: https://github.com/alex000kim/claude-code/blob/main/src/util...
That seems desirable? Like that's what commit messages are for. Describing the change. Much rather that than the m$ way of putting ads in commit messages
The commit message should complement the code. Ideally, what the code does should not need a separate description, but of course there can be exceptions. Usually, it's more interesting to capture in the commit message what is not in the code: the reason why this approach was chosen and not some other obvious one. Or describe what is missing, and why it isn't needed.
That sounds like design discussions best had in the issue/ticket itself, before you even start writing code. Then the commit message references the ticket and has a brief summary of the changes.
Writing and reading paragraphs of design discussion in a commit message is not something that seems common.
Ticket systems are quite ephemeral. I still have access to commit messages from the 90s (and I didn't work on the software at the time). I haven't been able to track the contents of the gnats bug tracker from those days.
And of course tickets can be private, so even if the data survived migration, you may not have access to it (principle of least privilege and all that).
Not really about design, but technical reasons why this solution came to be when it’s not that obvious. It’s not often needed. And when it does, it usually fits in a short paragraph.
What you're describing here is a design. The most important parts of a design are the decisions and their reasoning.
e.g. "we decided on tool/library pattern X over tool/library/pattern Y because Z" – that is a design, usually discussed outside (and before) a commit message.
You discuss these decisions with others, document the discussion and decision, and then you have a design and can start writing code.
Let me ask you this: suppose you have a task that needs to be done eventually, and you want to write down some ideas for it, but don't want to start coding right now. Where do you put those ideas? How do you link them to that specific task?
Unfortunately GitHub Copilot’s commit message generation feature is very human. It’s picked up some awful habits from lazy human devs. I almost always get some pointless “… to improve clarity” or “… for enhanced usability” at the end of the message.
VS Code has a setting that promises to change the prompt it uses to generate commit messages, but it mostly ignores my instructions, even very literal ones like “don’t use the words ‘enhance’ or ‘improve’”. And oddly having it set can sometimes result in Cyrillic characters showing up at the end of the message.
Ultimately I stopped using it, because editing the messages cost me more time than it saved.
Honestly the aggressive verbosity of github copilot is half the reason don't use its suggested comments. AI generated code comments follow an inverted-wadsworth-constant: Only the first 30% is useful.
As opposed to outputting debugging information, which I wouldnt be surprised if LLMs do output "debug" output blurbs which could include model specific information.
The human developer would just write what the code does, because the commit also contains an email address that identifies who wrote the commit. There's no reason to write:
> Commit f9205ab3 by dkenyser on 2026-3-31 at 16:05:
> Fixed the foobar bug by adding a baz flag - dkenyser
Because it already identified you in the commit description. The reason to add a signature to the message is that someone (or something) that isn't you is using your account, which seems like a bad idea.
Aside from merges that combine commits from many authors onto a production branch or release tag. I would personally not leave an agent to do that sort of work.
Which is funny given how many workplaces are requiring developers use AI, measuring their usage, and stack ranking them by how many tokens they burn. What I want is something that I can run my human-created work product through to fool my employer and its AI bean counters into thinking I used AI to make it.
I guess that would work until they started auditing your prompts. I suppose you could just have a background process on your workstation just sitting there Clauding away on the actual problem, while you do your development work, and then just throw away the LLM's output.
Undercover mode seems like a way to make contributions to OSS when they detect issues, without accidentally leaking that it was claude-mythos-gigabrain-100000B that figured out the issue
I'm amazed at how much of what my past employers would call trade secrets are just being shipped in the source. Including comments that just plainly state the whole business backstory of certain decisions. It's like they discarded all release harnesses and project tracking and just YOLO'd everything into the codebase itself.
Edit: Everyone is responding "comments are good" and I can't tell if any of you actually read TFA or not
> “BQ 2026-03-10: 1,279 sessions had 50+ consecutive failures (up to 3,272) in a single session, wasting ~250K API calls/day globally.”
This is just revealing operational details the agent doesn't need to know to set `MAX_CONSECUTIVE_AUTOCOMPACT_FAILURES = 3`
Agents and I apparently have a whole lot in common.
Only being half ironic with this. I generally find that people somehow magically manage to understand how to be materially helpful when the subject is a helpless LLM. Instead of pointing it to a random KB page, they give it context. They then shorten that context. They then interleave context as comments. They provide relevant details. They go out of their way to collect relevant details. Things they somehow don't do for their actual colleagues.
This only gets worse when the LLM captures all that information better than certain human colleagues somehow, rewarding the additional effort.
Right? It's infuriating. Nearly all of the agentic coding best practices are things that we should have just been doing all along, because it turns out humans function better too when given the proper context for their work. The only silver lining is that this is a colossal karmic retribution for the orgs that never gave a shit about this stuff until LLMs.
> If you're not using comments, you're doing agent coding wrong.
Comments are ultimately so you can understand stuff without having to read all the code. LLMs are great when you force them to read all code, and comments only serve to confuse. I'd say the opposite been true in my experience, if you're not forcing LLMs to not have any comments at all (and it can actually skip those, looking at you Gemini), you're doing agent coding wrong.
Meanwhile, some colleagues: "Code should have as little comments as possible, the code should explain itself." (conceptually not wholly wrong, but it can only explain HOW not WHY and even then often insufficiently) all while having barebones/empty README.md files more often than not. Fun times.
This is also a great way to ensure the documentation is up to date. It’s easier to fix the comment while you’re in the code just below it than to remember “ah yes I have to update docs/something.md because I modified src/foo/bar.ts”.
People moving docs out of code are absolutely foolish because no one is going to remember to update it consistently but the agent always updates comments in the line of sight consistently.
Agent is not going to know to look for a file to update unless instructed. Now your file is out of sync. Code comments keeping everything line of sight makes it easy and foolproof.
Tbh, I think distillation is happening both ways. And at this stage, "quality" is stagnating, the main edge is the tooling. The harness of CC seems to be the best so far, and I wonder if this leak would equalize the usability.
more likely, they would parse them out using simple regex, the whole point is they're there but not used. Distillation is becoming less common now however
I don't understand the part about undercover mode. How is this different from disabling claude attribution in commits (and optionally telling claude to act human?)
On that note, this article is also pretty obviously AI-generated and it's unfortunate the author didn't clean it up.
It's people overreacting, the purpose of it is simple, don't leak any codenames, project names, file names, etc when touching external / public facing code that you are maintaining using bleeding edge versions of Claude Code. It does read weird in that they want it to write as if a developer wrote a commit, but it might be to avoid it outputting debug information in a commit message.
The buddy feature the article mentions is planned for release tomorrow, as a sort of April Fools easter egg. It'll roll out gradually over the day for "sustained Twitter buzz" according to the source.
The pet you get is generated based off your account UUID, but the algorithm is right there in the source, and it's deterministic, so you can check ahead of time. Threw together a little app to help, not to brag but I got a legendary ghost https://claudebuddychecker.netlify.app/
Cute! Cactus for me. Nice animations too - looks like there were multiple of us asking Claude to reverse engineer the system. I did a slightly deeper dive here if you're interested, plus you can see all the options available: https://variety.is/posts/claude-code-buddies/
(I didn't think to include a UUID checker though - nice touch)
Neat! That's a great write up, cool to see others looking into it.
I do wonder if they're going to do anything with the stats and shinies bit. Seems like the main piece of code for buddies that's going to handle hatching them tomorrow is still missing (comments mention a missing /buddy/index file), so maybe it'll use them there.
Langchain is for model-agnostic composition. Claude Code only uses one interface to hoist its own models so zero need for an abstraction layer.
Langgraph is for multi-agent orchestration as state graphs. This isn't useful for Claude Code as there is no multi-agent chaining. It uses a single coordinator agent that spawns subagents on demand. Basically too dynamic to constrain to state graphs.
I'm not an supporter of blindly adopting the "langs" but langgraph is useful for deterministically reproducable orchestration. Let's say you have a particular data flow that takes an email sends it through an agent for keyword analysis the another agent for embedding then splits to two agents for sentiment analysis and translation - there is where you'd use langgraph in your service. Claude Code is a consumer tool, not production.
I see what you mean. Maybe in the cases where the steps are deterministic, it might be worth moving the coordination at the code layer instead of AI layer.
What's the value add over doing it with just Python code? I mean you can represent any logic in terms of graphs and states..
I've tried to use langchain. It seemed to force code into their way of doing things and was deeply opinionated about things that didn't matter like prompt templating. Maybe it's improved since then, but I've sort of used people who think langchain is good as a proxy for people who haven't used much ai?
Maybe it would be okay as a first filtering step, before doing actual sentiment analysis on the matches. That would at least eliminate obvious false positives (but of course still do nothing about false negatives).
Is this really the use-case? I imagine the regex is good for a dashboard. You can collect matches per 1000 prompts or something like that, and see if the number grows or declines over time. If you miss some negative sentiment it shouldn't matter unless the use of that specific word doesn't correlate over time with other negative words and is also popular enough to have an impact on the metric.
The exploit is a postinstall hook, so CC users would be unaffected. Claude Code itself is most likely built with bun and not npm, so the CC developers would also be immune.
> The obvious concern, raised repeatedly in the HN thread: this means AI-authored commits and PRs from Anthropic employees in open source projects will have no indication that an AI wrote them. It’s one thing to hide internal codenames. It’s another to have the AI actively pretend to be human.
I don’t get it. What does this mean? I can use Claude code now without anyone knowing it is Claude code.
it's written to _actively_ avoid any signs of AI generated code when "in a PUBLIC/OPEN-SOURCE repository".
Also, it's not about you. Undercover mode only activates for Anthropic employees (it's gated on USER_TYPE === 'ant', which is a build-time flag baked into internal builds).
There are probably different reasons for different people. I can definitely see the angle that trying to specifically pretend to not be AI when contributing to open source could be seen as a bad thing due to the open source supply chain attacks, some AI-driven, that we've been having, not to mention the AI-slop PR spam.
But, I also get Anthropic's side that when they're contributing they don't want their internals leaked. If it had been left at that, that's fine, but having it pretend like it's not AI at all rubs me a little bit the wrong way. Why try to hide it?
>There are probably different reasons for different people. I can definitely see the angle that trying to specifically pretend to not be AI when contributing to open source could be seen as a bad thing due to the open source supply chain attacks, some AI-driven, that we've been having, not to mention the AI-slop PR spam.
But none of the other agents advertise that the commit was done by an agent. Like Codex. Your panic should apply equally to already existing agents like Codex no?
Can someone clarify how the signing can't be spoofed (or can it)? If we have the source, can't we just use the key to now sign requests from other clients and pretend they're coming from CC itself?
Ah yes, the API will accept requests that doesn’t include the client attestation (or the fingerprint from src/utils/fingerprint.ts. At least it did a couple of weeks back.
They are most likely using these as post-fact indicators and have automation they kicks in after a threshold is reached.
Now that the indicators have leaked, they will most likely be rotated.
Anyone else have CI checks that source map files are missing from the build folder? Another trick is to grep the build folder for several function/variable names that you expect to be minified away.
> Anti-distillation: injecting fake tools to poison copycats
Does this mean `huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled` is unusable? Had anyone seen fake tool calls working with this model?
The feature flag names alone are more revealing than the code. KAIROS, the anti-distillation flags, model codenames those are product strategy decisions that competitors can now plan around. You can refactor code in a week. You can't un-leak a roadmap.
>This was the most-discussed finding in the HN thread. The general reaction: an LLM company using regexes for sentiment analysis is peak irony.
>Is it ironic? Sure. Is it also probably faster and cheaper than running an LLM inference just to figure out if a user is swearing at the tool? Also yes. Sometimes a regex is the right tool.
I'm reading an LLM written write up on an LLM tool that just summarizes HN comments.
I'm so tired man, what the hell are we doing here.
It's possible, but Anthropic employees regularly boast (!) that Claude Code is itself almost entirely vibe-coded (which certainly seems true, based on the generally-low quality of the code in this leak), so it wouldn't at all surprise me to have that blow up twice in the same week. Probably it might happen with accelerating frequency as the codebase gets more and more unmanageable.
I'm normally suspicious but honestly they've been so massively supply-constrained that I don't think it really benefits them much. They're not worried about getting enough demand for the new models; they're worrying about keeping up with it.
Granted, there's a small counterargument for mythos which is that it's probably going to be API-only not subscription
Something I’ve been thinking about, somewhat related but also tangential to this topic:
The more code gets generated by AI, won’t that mean taking source code from a company becomes legal? Isn’t it true that works created with generative AI can’t be copyrighted?
I wonder if large companies have throught of this risk. Once a company’s product source code reaches a certain percentage of AI generation it no longer has copyright. Any employee with access can just take it and sell it to someone else, legally, right?
Well, as a general rule, I don't do business with people who lie to me.
You've got a business, and you sent me junk mail, but you made it look like some official government thing to get me to open it? I'm done, just because you lied on the envelope. I don't care how badly I need your service. There's a dozen other places that can provide it; I'll pick one of them rather than you, because you've shown yourself to be dishonest right out of the gate.
Same thing with an AI (or a business that creates an AI). You're willing to lie about who you are (or have your tool do so)? What else are you willing to lie to me about? I don't have time in my life for that. I'm out right here.
Out of curiosity, given two code submissions that are completely identical—one written solely by a human and one assisted by AI—why should its provenance make any difference to you? Is it like fine art, where it’s important that Picasso’s hand drew it? Or is it like an instruction manual, where the author is unimportant?
Similarly, would you consider it to be dishonest if my human colleague reviewed and made changes to my code, but I didn’t explicitly credit them?
As an attorney, I know copyright law. (This is not legal advice.) There's nothing about copyright law that says you have to credit an AI coding agent for contributing to your work. The person receiving the code has to perform their due diligence in any case to determine whether the author owns it or has permission from the owner to contribute it.
Why does the provenance make any difference? Let me increase your options. Option 1: You completely hand-wrote it. Option 2: You were assisted by an AI, but you carefully reviewed it. Option 3: You were assisted by an AI (or the AI wrote the whole thing), and you just said, "looks good, YOLO".
Even if the code is line-for-line identical, the difference is in how much trust I am willing to give the code. If I have to work in the neighborhood of that code, I need to know what degree of skepticism I should be viewing it with.
That's the thing. As someone evaluating pull requests, should you trust the code based on its provenance, or should you trust it based on its content? Automated testing can validate code, but it can't validate people.
ISTM the most efficient and objective solution is to invest in AI more on both sides of the fence.
You are spamming the whole fucking thread with the same nonsense. It is instructed to hide that the PR was made via Claude Code. I don't know why people who are so AI forward like yourself have such a problem with telling people that they use AI for coding/writing, it's a weirdly insecure look.
what's insecure about it? if it is up to the institution to make that decision - you can still do it. Claude is not stopping you from making that decision