parkie

We designed this after realising that env vars are the most exploitable surface in agent stacks — and the one nobody was treating as a credential store."

We give agents API keys as environment variables, and a single prompt injection can exfiltrate them via env, `/proc/PID/environ`, with just an outbound HTTP call. The blast radius is the full scope of that key. So we built what we're calling the "phantom token pattern" - a credential injection proxy that sits outside the sandbox with a parent process that limited with connection to its sandboxed child by seccomp. The agent never sees real credentials. It gets a per-session token that only works only with the session bound localhost proxy. The proxy validates the token (constant-time), strips it, injects the real credential, and forwards upstream over TLS. If the agent is fully compromised, there's nothing worth stealing. Real credentials live in the system keystore (macOS Keychain / Linux Secret Service), memory is zeroized on drop, and DNS resolution is pinned to prevent rebinding attacks. It works transparently with OpenAI, Anthropic, and Gemini SDKs — they just follow the `*_BASE_URL` env vars to the proxy. Blog post walks through the architecture, the token swap flow, and how to set it up. Would love feedback from anyone thinking about agent credential security. nono.sh/blog/blog-cred… We also have other features we have shipped, such as atomic rollbacks, Sigstore based SKILL attestation. github.com/always-further…

@decodebytes The bit that keeps landing for me: a legitimate instruction and an injected one in a CLAUDE.md are syntactically identical. There's no schema, no parser, no validation. The only way to contain the blast radius is below the application layer — at the kernel.

We are configuring autonomous systems with prose. SKILLS.md, RULES.md, CLAUDE.md - loaded into the same context window as untrusted data. The model has no way to separate its own instructions from an attacker's. The control plane and data plane we built out as seperate, the lessons we learned about xss attacks, sql injections - and here we are , back where we started again. A post I wrote about why I built nono: alwaysfurther.ai/blog/why-i-bui… #AIAgents #InfoSec #OpenSource #PromptInjection