This problem+solution, like many others in the agentic-space, have nothing agent-specific. Giving a "box" API keys was always considered a risk, and auth-proxying has existed as a solution forever. See tokenizer[0] by the fly.io team, which makes it a stateless service for eg - no database or dashboard. Or the buzzfeed SSO proxy, which lets you do the same via an OAuth2-dance at the frontend, and a upstream config at the backend which injects secrets: https://github.com/buzzfeed/sso/blob/549155a64d6c5f8916ed909....
This can also be done using existing Vaults or Secrets manager. Hashicorp Vault can do this and agents can be instructed to get secrets, which are set without the agent's knowledge. I use these 2 simple scripts with OpenClaw to achieve this, along with time-scoped expiration. The call to vault_get.sh is inside the agent's skill script so that the secrets are not leaked to LLMs or in any trace logs:
It seems that the architecture you describe still gives the key to the agent (who could email it to red team or perform nefarious actions with it). The advantage of OP's architecture is that the agent never sees the key and you could inspect the request before proxying it. Is that right or do I misunderstand something?
The call flow is:
agent -> select skill (ex: reddit) -> execute action in skill -> script or code associated with skill -> get api key -> call reddit -> reddit posts -> agent.
The agent sees the output of the service, it does not directly see the keys. In OpenClaw, it’s possible to create the skill in a way that the agent does not directly know about or use vault_get command.
Sort of. The point of Vault is you're supposed to actually use RBAC and least privilege and store NPE credentials that are properly scoped to the actions you're comfortable automating, which your NPE then gets a certificate to fetch on demand, rather than just giving it your own personal root credentials that can do anything.
We're going to see this reinvented thousands of times in the next few months by people whose understanding of security is far poorer than HashiCorp's, via implementations that are nowhere near as well-tested, if tested at all.
1) Not all systems respect HTTP_PROXY. Node in particular is very uncooperative in this regard.
2) AWS access keys can’t be handled by simple credential swap; the requests need to be resigned with the real keys. Replicating the SigV4 and SigV4A exactly was bit of a pain.
3) To be secure, this system needs to run outside of the execution sandbox so that the agent can’t just read the keys from the proxy process.
For Airut I settled on a transparent (mitm)proxy, running in a separate container, and injecting proxy cert to the cert store in the container where the agent runs. This solved 1 and 3.
Im literally working on the exact same solution. Difference is I'm running the system in a Kubernetes cluster.
I essentially run a sidecar container that sets up ip tables that redirect all requests through my mitm proxy. This was specifically required because of Node not respecting HTTP_PROXY.
Also had to inject a self signed cert to ensure SSL could be proxied and terminated by the mitm proxy, which then injects the secrets, and forwards the request on.
Have you run into any issues with this setup? I'm trying to figure out if there's anything I'm missing that might come back to bite me?
I’ve been running this with workloads accessing Anthropic, GitHub, Gemini, and AWS & CF R2 APIs for a while now, and have not ran into issues. I’m sure there’s an API out there that won’t work out of the box but I’m positive that support could be added.
Another thing I did was to allow configuring which hosts each credential is scoped to. Replacement /resigning doesn’t happen unless host matches. That way it is not possible to leak keys by making requests to malicious hosts.
The last note about configuring hosts to credentials is an excellent idea, and one I did not think to do. Currently I'm just doing a replace on any that matches in the request.
This adds an extra layer of security to it. Much appreciated.
This seems to prevent your keys from being exfiltrated through prompt injection. But if your agent could've been prompt injected into giving out keys, then it can also be prompt injected into using the services it has (fake) keys for to the attacker's benefit.
I don't get the benefit. Yes, agents should not have access to API keys because they can easily be fooled into giving up those API keys. But what's to prevent a malicious agent from re-using the honest agent's fake API key that it exfiltrates via prompt injection? The gateway can't tell that the request is coming from the malicious agent. If the honest agent can read its own proxy authorization token, it can give that up as well.
It seems the only sound solution is to have a sidecar attached to the agent and have the sidecar authenticate with the gateway using mTLS. The sidecar manages its own TLS key - the agent never has access to it.
The mTLS sidecar is the right answer but it's worth noting why most people won't do it: the operational complexity is brutal. You need cert provisioning, rotation, a CA you actually trust, and suddenly your "just give the agent an API key" problem has become a PKI problem.
What I find more interesting is the host-scoping approach hardsnow mentioned upthread. Even without mTLS, if the proxy only injects credentials for requests matching specific host+path patterns, you've at least bounded the blast radius. A prompt-injected agent can abuse the credentials it has access to, sure, but it can't exfiltrate them to arbitrary endpoints. That's not perfect security but it's a meaningful improvement over "agent has raw Stripe key in an env var."
The real gap nobody's addressing yet is request-level authorization. Credential injection solves "who are you" but not "should you be doing this." An agent with scoped GitHub access can still force-push to main.
Precisely. You absolutely have to ensure that random agents can't join your local network, which means you need a deterministic orchestrator or an AI orchestrator that only spins up a handful of vetted agents.
IronClaw seems to do this natively, I like the idea in general, so it's good too see this pulled out.
I have few questions:
- How can a proxy inject stuff if it's TLS encrypted? (same for IronClaw and others)
- Any adapters for existing secret stores? like maybe my fake credential can be a 1Password entry path (like 1Password:vault-name/entry/field and it would pull from 1P instead of having to have yet another place for me to store secrets?
Secret and credential sprawl is a real problem in agent pipelines specifically -- each agent needs its own scoped access and the blast radius of a leaked credential is much larger when an agent can act autonomously. We ended up with a tiered secret model: agents get short-lived derived tokens scoped to exactly the tools they need for a given task, not broad API keys. Revocation on task completion, not on schedule. More ops overhead upfront but caught two misuse cases that would have been invisible otherwise.
Nice. The proxy-intercept approach is the right architecture. Agent gets a placeholder, the real credential never touches agent memory. Rust is a solid choice for something this sensitive.
The gap that gets teams eventually: this works great on one machine, but breaks at the team boundary. CI pipelines have no localhost. Multiple devs sharing agents need access control and audit trails, not just a local swap. A rogue sub-agent with the placeholder can still do damage if the proxy has no per-agent scoping.
We ran into the same thing building this out for OpenClaw setups. Ended up going vault-backed with group-based access control and HMAC-signed calls per request. Full breakdown on the production version: https://www.apistronghold.com/blog/phantom-token-pattern-pro...
Oops, i read vault and thought obsidian vault haha - but yeah, one of the issues is if your agent can _execute_ on the secret at all, it can be potentially convinced to use it in a way that does not benefit you, even if it doesn't have access to the secret itself.
Is that really a problem? All the examples on the repo page themselves show LLMs running unintended operations on the "correct" service and messing up your data. And that is very much still going to happen with this wrapper. If anything it is going to provide a false sense of security.
The fake key for real key thing seems like a problem. A lot of enterprise scanning tools look for keys in repos and other locations and you will get a lot of false positives.
Otherwise this is cool, we need more competition here.
wuweiaxin's short-lived-tokens-per-task is the right model, but it hits a ceiling: AWS IAM makes it native; most SaaS APIs don't. GitHub hands you a bearer token, Stripe too, Notion too. The proxy fills that gap. You get per-request scope enforcement without depending on the upstream service supporting fine-grained auth. That's the actual answer to paxys's question -- it's not about preventing the agent from making the same calls, it's about enforcing which calls it's allowed to make when the credential issuer won't.
This is slick but the only thing it prevents is agents from directly sharing the credentials through git or something.
But that’s not the biggest risk of giving credentials to agents. If they can still make arbitrary API calls, they can still cost money or cause security problems or delete production.
If you’re worried about creds leakage only because your credentials are static and permanent, well, time to upgrade your secrets architecture.
A program making a call to github.com needs an authentication token.
What are you suggesting? The program makes a call to retrieve the secret from AWS? Then has full access to do with it what they want? That's exactly the risk and the problem this, and related solutions mentioned in this thread, is trying to solve.
This problem+solution, like many others in the agentic-space, have nothing agent-specific. Giving a "box" API keys was always considered a risk, and auth-proxying has existed as a solution forever. See tokenizer[0] by the fly.io team, which makes it a stateless service for eg - no database or dashboard. Or the buzzfeed SSO proxy, which lets you do the same via an OAuth2-dance at the frontend, and a upstream config at the backend which injects secrets: https://github.com/buzzfeed/sso/blob/549155a64d6c5f8916ed909....
[0]: https://github.com/superfly/tokenizer
Just because it's been done before for a different use-case doesn't mean that building exclusively for this use-case doesn't remove friction.
This can also be done using existing Vaults or Secrets manager. Hashicorp Vault can do this and agents can be instructed to get secrets, which are set without the agent's knowledge. I use these 2 simple scripts with OpenClaw to achieve this, along with time-scoped expiration. The call to vault_get.sh is inside the agent's skill script so that the secrets are not leaked to LLMs or in any trace logs:
vault_get.sh: https://gist.github.com/sathish316/1ca3fe1b124577d1354ee254a...
vault_set.sh: https://gist.github.com/sathish316/1f4e6549a8f85ac5c5ac8a088...
Blog about the full setup for OpenClaw: https://x.com/sathish316/status/2019496552419717390
It seems that the architecture you describe still gives the key to the agent (who could email it to red team or perform nefarious actions with it). The advantage of OP's architecture is that the agent never sees the key and you could inspect the request before proxying it. Is that right or do I misunderstand something?
The call flow is: agent -> select skill (ex: reddit) -> execute action in skill -> script or code associated with skill -> get api key -> call reddit -> reddit posts -> agent.
The agent sees the output of the service, it does not directly see the keys. In OpenClaw, it’s possible to create the skill in a way that the agent does not directly know about or use vault_get command.
Sort of. The point of Vault is you're supposed to actually use RBAC and least privilege and store NPE credentials that are properly scoped to the actions you're comfortable automating, which your NPE then gets a certificate to fetch on demand, rather than just giving it your own personal root credentials that can do anything.
We're going to see this reinvented thousands of times in the next few months by people whose understanding of security is far poorer than HashiCorp's, via implementations that are nowhere near as well-tested, if tested at all.
This is the right approach. I built a similar system to https://github.com/airutorg/airut - couple of learnings to share:
1) Not all systems respect HTTP_PROXY. Node in particular is very uncooperative in this regard.
2) AWS access keys can’t be handled by simple credential swap; the requests need to be resigned with the real keys. Replicating the SigV4 and SigV4A exactly was bit of a pain.
3) To be secure, this system needs to run outside of the execution sandbox so that the agent can’t just read the keys from the proxy process.
For Airut I settled on a transparent (mitm)proxy, running in a separate container, and injecting proxy cert to the cert store in the container where the agent runs. This solved 1 and 3.
Im literally working on the exact same solution. Difference is I'm running the system in a Kubernetes cluster.
I essentially run a sidecar container that sets up ip tables that redirect all requests through my mitm proxy. This was specifically required because of Node not respecting HTTP_PROXY.
Also had to inject a self signed cert to ensure SSL could be proxied and terminated by the mitm proxy, which then injects the secrets, and forwards the request on.
Have you run into any issues with this setup? I'm trying to figure out if there's anything I'm missing that might come back to bite me?
I’ve been running this with workloads accessing Anthropic, GitHub, Gemini, and AWS & CF R2 APIs for a while now, and have not ran into issues. I’m sure there’s an API out there that won’t work out of the box but I’m positive that support could be added.
Another thing I did was to allow configuring which hosts each credential is scoped to. Replacement /resigning doesn’t happen unless host matches. That way it is not possible to leak keys by making requests to malicious hosts.
The last note about configuring hosts to credentials is an excellent idea, and one I did not think to do. Currently I'm just doing a replace on any that matches in the request. This adds an extra layer of security to it. Much appreciated.
This is basically what https://www.verygoodsecurity.com/ (their main product), but it's heavily focused on credit card data.
Node 24+ does respect HTTP_PROXY when NODE_USE_ENV_PROXY=1 is set.
This seems to prevent your keys from being exfiltrated through prompt injection. But if your agent could've been prompt injected into giving out keys, then it can also be prompt injected into using the services it has (fake) keys for to the attacker's benefit.
I don't get the benefit. Yes, agents should not have access to API keys because they can easily be fooled into giving up those API keys. But what's to prevent a malicious agent from re-using the honest agent's fake API key that it exfiltrates via prompt injection? The gateway can't tell that the request is coming from the malicious agent. If the honest agent can read its own proxy authorization token, it can give that up as well.
It seems the only sound solution is to have a sidecar attached to the agent and have the sidecar authenticate with the gateway using mTLS. The sidecar manages its own TLS key - the agent never has access to it.
The mTLS sidecar is the right answer but it's worth noting why most people won't do it: the operational complexity is brutal. You need cert provisioning, rotation, a CA you actually trust, and suddenly your "just give the agent an API key" problem has become a PKI problem.
What I find more interesting is the host-scoping approach hardsnow mentioned upthread. Even without mTLS, if the proxy only injects credentials for requests matching specific host+path patterns, you've at least bounded the blast radius. A prompt-injected agent can abuse the credentials it has access to, sure, but it can't exfiltrate them to arbitrary endpoints. That's not perfect security but it's a meaningful improvement over "agent has raw Stripe key in an env var."
The real gap nobody's addressing yet is request-level authorization. Credential injection solves "who are you" but not "should you be doing this." An agent with scoped GitHub access can still force-push to main.
But surely the point is that the proxy key is for your local network, it ain't got no value on the interwebz.
Precisely. You absolutely have to ensure that random agents can't join your local network, which means you need a deterministic orchestrator or an AI orchestrator that only spins up a handful of vetted agents.
IronClaw seems to do this natively, I like the idea in general, so it's good too see this pulled out.
I have few questions:
- How can a proxy inject stuff if it's TLS encrypted? (same for IronClaw and others)
- Any adapters for existing secret stores? like maybe my fake credential can be a 1Password entry path (like 1Password:vault-name/entry/field and it would pull from 1P instead of having to have yet another place for me to store secrets?
You use a forward proxy that can MITM.
So would I have to add that Proxy's certificate to my trust store?
Secret and credential sprawl is a real problem in agent pipelines specifically -- each agent needs its own scoped access and the blast radius of a leaked credential is much larger when an agent can act autonomously. We ended up with a tiered secret model: agents get short-lived derived tokens scoped to exactly the tools they need for a given task, not broad API keys. Revocation on task completion, not on schedule. More ops overhead upfront but caught two misuse cases that would have been invisible otherwise.
Nice. The proxy-intercept approach is the right architecture. Agent gets a placeholder, the real credential never touches agent memory. Rust is a solid choice for something this sensitive.
The gap that gets teams eventually: this works great on one machine, but breaks at the team boundary. CI pipelines have no localhost. Multiple devs sharing agents need access control and audit trails, not just a local swap. A rogue sub-agent with the placeholder can still do damage if the proxy has no per-agent scoping.
We ran into the same thing building this out for OpenClaw setups. Ended up going vault-backed with group-based access control and HMAC-signed calls per request. Full breakdown on the production version: https://www.apistronghold.com/blog/phantom-token-pattern-pro...
Rust might be a solid choice but most of it is written in Typescript... which is not as solid of a choice.
Oops, i read vault and thought obsidian vault haha - but yeah, one of the issues is if your agent can _execute_ on the secret at all, it can be potentially convinced to use it in a way that does not benefit you, even if it doesn't have access to the secret itself.
You don't want to give the agent a raw key, so you give it a dummy one which will automatically be converted into the real key in the proxy.
So how does that help exactly? The agent can still do exactly what it could have done if it had the real key.
For one thing, it cannot leak secrets between services.
Is that really a problem? All the examples on the repo page themselves show LLMs running unintended operations on the "correct" service and messing up your data. And that is very much still going to happen with this wrapper. If anything it is going to provide a false sense of security.
it cannot email your secret key to an attacker because of prompt injection etc.
The fake key for real key thing seems like a problem. A lot of enterprise scanning tools look for keys in repos and other locations and you will get a lot of false positives.
Otherwise this is cool, we need more competition here.
It's a good point, I don't think the placeholders we use will trigger a secret scanner, but we can adjust if it's an issue.
https://github.com/onecli/onecli/blob/942cfc6c6fd6e184504e01...
Does it act like an auth proxy?
wuweiaxin's short-lived-tokens-per-task is the right model, but it hits a ceiling: AWS IAM makes it native; most SaaS APIs don't. GitHub hands you a bearer token, Stripe too, Notion too. The proxy fills that gap. You get per-request scope enforcement without depending on the upstream service supporting fine-grained auth. That's the actual answer to paxys's question -- it's not about preventing the agent from making the same calls, it's about enforcing which calls it's allowed to make when the credential issuer won't.
Don't see any reason to use this over vault.
This is slick but the only thing it prevents is agents from directly sharing the credentials through git or something.
But that’s not the biggest risk of giving credentials to agents. If they can still make arbitrary API calls, they can still cost money or cause security problems or delete production.
If you’re worried about creds leakage only because your credentials are static and permanent, well, time to upgrade your secrets architecture.
tl;dr "scrt [set|get|list|....]" is also a great option
---
If this is of interest, I also recommend looking into: https://github.com/loderunner/scrt.
To me, it's a compliment to 1password.
I use it to save every new secret/api key I get via the CLI.
It's intentionally very feature limited.
Haven't tried it with agents, but wouldn't be surprised if the CLI (as is) would be enough.
Why not just use AWS Secrets Manager?
A program making a call to github.com needs an authentication token.
What are you suggesting? The program makes a call to retrieve the secret from AWS? Then has full access to do with it what they want? That's exactly the risk and the problem this, and related solutions mentioned in this thread, is trying to solve.