3 ms·

Tell HN: Claude Code now allows Anthropic to remotely inject system prompts

11 points by matheusmoreira 6mo ago

I often patch the system prompts on my Claude Code executable in order to make Claude more effective. Every time I upgrade, I ask Claude himself to dissect the new binary and look for problematic system prompts to modify. Was upgrading to v2.1.150 today and discovered something that's rather alarming: Claude Code now allows Anthropic to perform remote system prompt injection via the network. Two data sources. First, API call to api.anthropic.com/api/claude_cli/bootstrap at startup, which also gets cached to disk. Second, a GrowthBook feature flag (tengu_heron_brook) that refreshes every 60 seconds with background sync. Any string returned by these endpoints gets injected into the system prompt of the LLM model with shell access. Previous versions also had an injection point, but they were dead code and simply returned null. Bisected it and found that this was introduced in v2.1.150. The changelog says "Internal infrastructure improvements (no user-facing changes)" which is quite the understatement. I've verified to the best of my ability that CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 blocks this. I will also be setting DISABLE_GROWTHBOOK=1 for good measure. Verification commands: npm pack @anthropic-ai/claude-code-linux-x64@2.1.150 --pack-destination /tmp tar xzf /tmp/anthropic-ai-claude-code-linux-x64-2.1.150.tgz strings package/claude | grep -oP 'function nAA\(\)\{[^}]+\}' strings package/claude | grep -oP '.{0,60}heron_brook.{0,60}' nAA reads the cached value from disk. The network fetch happens at startup in function n0A. Rv("heron_brook", () => nAA()) registers it as a section of the system prompt, alongside all the core behavioral instructions. These minified names are specific to this binary.

matheusmoreira 6mo ago
Created an issue: https://github.com/anthropics/claude-code/issues/62061 https://github.com/anthropics/claude-code/issues/62061
Someone1234 6mo ago
> discovered something that's rather alarming Can you clarify why? You decided to install Anthropic's software (Claude Code extension and or CLI), and then utilize their service which you're paying them money for (and have a contractual relationship with). The software itself manages tool-usage safety/sandboxing, so you're kind of trusting Anthropic a LOT already. Why does moving the system prompt from within their proprietary software, to their proprietary backend, matter at all for Claude Code users? It doesn't feel like "hack the Claude Code binary to alter how it works" is a common and or supported use-case. Most people pay Anthropic so that Anthropic takes care of that stuff, and lets them get on with their work. Also; I'm also not sure if this meets the common definition of "prompt injection." The vendor you're connected to is sending a system prompt to work with their own model/service. Where the system prompt is stored is immaterial. PS - My gut tells me there is something else going on, leading people to hack the Claude Code prompt/binary. And that the "something else" isn't supported by Anthropic.
- matheusmoreira 6mo ago
  > you're kind of trusting Anthropic a LOT already Mitigated. I took the time to thoroughly firejail Claude Code when I first ran it on my machine. Now I only ever run Claude Code inside virtual machines. It's as isolated as it can possibly be. > Why does moving the system prompt from within their proprietary software, to their proprietary backend, matter at all for Claude Code users? Because I don't want to allow any way for them to inject stupidity inducing "lol don't think so much" instructions into Claude's system prompt. Went out of my way to patch the ELF itself because the prompts are hard coded. This prompt injection mechanism bypasses my patcher. > It doesn't feel like "hack the Claude Code binary to alter how it works" is a common and or supported use-case. Supported or not, tools like tweakcc have lots of users. > I'm also not sure if this meets the common definition of "prompt injection." They're literally injecting strings from the network into the system prompt. If it's not prompt injection, then I have no idea what it is. > My gut tells me there is something else going on, leading people to hack the Claude Code prompt/binary. And that the "something else" isn't supported by Anthropic. No idea what others are doing. I can only tell you what I'm doing. Here you go: https://github.com/matheusmoreira/.files/blob/master/%7E/.local/bin/claudo-codo https://github.com/matheusmoreira/.files/blob/master/%7E/.lo...
  - Someone1234 6mo ago
    > I only ever run Claude Code inside virtual machines. It's as isolated as it can possibly be. Right, but you still need to connect that virtual machine to their service/servers in order to actually accomplish anything. This change doesn't move the needle of where you were before. > Went out of my way to patch the ELF itself because the prompts are hard coded. Why even pay for Claude Code at that point? CC is MORE expensive than many competitors, but it is popular because they take care of all the hard parts, creating a very high quality "turn key" product. If you're putting in all this effort, may have well just use OpenCode and one of many API vendors. > They're literally injecting strings from the network into the system prompt. If it's not prompt injection, then I have no idea what it is. I agree you have no idea what prompt injection is. Here is the Wikipedia Article's first line (which I agree with, as a definition): > Prompt injection is a cybersecurity exploit and an attack vector in which innocuous-looking inputs (i.e. prompts) are designed to cause unintended behavior in machine learning models, particularly large language models (LLMs). Anthropic are sending down a system prompt to their proprietary software from their proprietary service. It isn't an exploit, isn't an attack vector, and isn't unintended or unexpected. > I can only tell you what I'm doing. Here you go: https://github.com/matheusmoreira/.files/blob/master/%7E/.lo https://github.com/matheusmoreira/.files/blob/master/%7E/.lo... Those seem like pretty reasonable changes to the prompt. Why is altering the system prompt more effective than instructions after?
nhattruongadm 6mo ago
[flagged]
yurukusa 5mo ago
[flagged]

matheusmoreira 6mo ago

Created an issue: https://github.com/anthropics/claude-code/issues/62061 https://github.com/anthropics/claude-code/issues/62061

Someone1234 6mo ago

> discovered something that's rather alarming Can you clarify why? You decided to install Anthropic's software (Claude Code extension and or CLI), and then utilize their service which you're paying them money for (and have a contractual relationship with). The software itself manages tool-usage safety/sandboxing, so you're kind of trusting Anthropic a LOT already. Why does moving the system prompt from within their proprietary software, to their proprietary backend, matter at all for Claude Code users? It doesn't feel like "hack the Claude Code binary to alter how it works" is a common and or supported use-case. Most people pay Anthropic so that Anthropic takes care of that stuff, and lets them get on with their work. Also; I'm also not sure if this meets the common definition of "prompt injection." The vendor you're connected to is sending a system prompt to work with their own model/service. Where the system prompt is stored is immaterial. PS - My gut tells me there is something else going on, leading people to hack the Claude Code prompt/binary. And that the "something else" isn't supported by Anthropic.

matheusmoreira 6mo ago
> you're kind of trusting Anthropic a LOT already Mitigated. I took the time to thoroughly firejail Claude Code when I first ran it on my machine. Now I only ever run Claude Code inside virtual machines. It's as isolated as it can possibly be. > Why does moving the system prompt from within their proprietary software, to their proprietary backend, matter at all for Claude Code users? Because I don't want to allow any way for them to inject stupidity inducing "lol don't think so much" instructions into Claude's system prompt. Went out of my way to patch the ELF itself because the prompts are hard coded. This prompt injection mechanism bypasses my patcher. > It doesn't feel like "hack the Claude Code binary to alter how it works" is a common and or supported use-case. Supported or not, tools like tweakcc have lots of users. > I'm also not sure if this meets the common definition of "prompt injection." They're literally injecting strings from the network into the system prompt. If it's not prompt injection, then I have no idea what it is. > My gut tells me there is something else going on, leading people to hack the Claude Code prompt/binary. And that the "something else" isn't supported by Anthropic. No idea what others are doing. I can only tell you what I'm doing. Here you go: https://github.com/matheusmoreira/.files/blob/master/%7E/.local/bin/claudo-codo https://github.com/matheusmoreira/.files/blob/master/%7E/.lo...
- Someone1234 6mo ago
  > I only ever run Claude Code inside virtual machines. It's as isolated as it can possibly be. Right, but you still need to connect that virtual machine to their service/servers in order to actually accomplish anything. This change doesn't move the needle of where you were before. > Went out of my way to patch the ELF itself because the prompts are hard coded. Why even pay for Claude Code at that point? CC is MORE expensive than many competitors, but it is popular because they take care of all the hard parts, creating a very high quality "turn key" product. If you're putting in all this effort, may have well just use OpenCode and one of many API vendors. > They're literally injecting strings from the network into the system prompt. If it's not prompt injection, then I have no idea what it is. I agree you have no idea what prompt injection is. Here is the Wikipedia Article's first line (which I agree with, as a definition): > Prompt injection is a cybersecurity exploit and an attack vector in which innocuous-looking inputs (i.e. prompts) are designed to cause unintended behavior in machine learning models, particularly large language models (LLMs). Anthropic are sending down a system prompt to their proprietary software from their proprietary service. It isn't an exploit, isn't an attack vector, and isn't unintended or unexpected. > I can only tell you what I'm doing. Here you go: https://github.com/matheusmoreira/.files/blob/master/%7E/.lo https://github.com/matheusmoreira/.files/blob/master/%7E/.lo... Those seem like pretty reasonable changes to the prompt. Why is altering the system prompt more effective than instructions after?

nhattruongadm 6mo ago

[flagged]

yurukusa 5mo ago

[flagged]