Standard OS Keychains have a massive security gap: any process running under your user context can query them.
A compromised npm package or python dependency running in the background can query the keychain and get your secrets.
Process-level validation was non-negotiable.
Built AgentSecrets because we got tired of AI agents touching raw credentials.
80+ authenticated calls this week.
0 credential exposures.
Agents pass key names only.
AgentSecrets injects credentials locally at the transport layer.
The coding assistant reading your repository right now can also read your environment variables and config files.
Stop using local plaintext files. Move your environment to the OS keychain and gate it behind process-level verification.
https://t.co/Url1C5MhIA
API keys don’t just leak in outbound requests. If an API response reflects the key, it enters the LLM’s context window.
AgentSecrets actively scans inbound responses and redacts reflected secrets before they reach the agent.
An agent with access to database tools can retrieve data. But where does it send it?
If a prompt injection tells the agent to upload the records to a foreign server, a simple network allowlist blocks it.
Domain boundaries are your final line of defense.
@theozero@pk_iv True, they all are connected in one way or the other. In the ai era redentials have stopped being just about access and secure storage, theirnuse must also be guarded
zero-knowledge gets you halfway.
the agent doesn't hold the credential. But what stops a hijacked agent from making calls
it was never supposed to make?
something we've been building.
The agent receives the error without the key. Your logs, debug consoles, and LLM context windows stay clean.
We've had 11 of such scenarios live according to our metrics.
How does AgentSecrets protect against ASI06: Memory & Context Poisoning?
When an upstream API reflects a credential back in an error payload, the agent's context is compromised. 👇
Most agents can read everything inside your .env file. So we built an AI secret infrastructure where credentials can be used without ever being exposed to the model.
Check out https://t.co/gcBEqzg3pv
@_agentsecrets_
AgentSecrets is not a secret manager.
It’s secrets infrastructure for AI agents.
Secret managers focus on storage and access. AgentSecrets focuses on usage.
Cuz the real problem isn’t just “can the agent access the secret?”
It’s “what can the agent do with it after access?”
API keys don’t just leak in outbound requests. If an API response reflects the key, it enters the LLM’s context window.
AgentSecrets actively scans inbound responses and redacts reflected secrets before they reach the agent.
You cannot secure an agent using prompts. Prompt engineering is a soft boundary; a clever user will always bypass it.
Security must be enforced by the network and operating system layers. Keep prompts for reasoning, use architecture for protection.