Skip to content

llm-shield-proxy 1.7.0

Choose a tag to compare

@ninadphalak ninadphalak released this 25 Oct 14:14
ca66748

llm-shield-proxy 1.7.0. Minor release: request-redaction and credential-routing fixes, with behaviour changes an operator should know about.

What is in it since 1.6.11

  • #208 Every value inside a chat message is scanned, whatever its shape. Real shapes that reached the provider unredacted:

    • OpenAI custom tool calls (custom.input);
    • Anthropic server_tool_use and mcp_tool_use input;
    • and any value of an unexpected type under a key the walk owns.

    Gemini's thought_signature, media payloads and file ids still go out unchanged.

  • #211

    • Client-picked hosts. A host the caller picks with X-Upstream-Base-Url or X-Shield-Fallback-Url gets a proxy-held key only if it is the configured upstream host or a known provider's API host, and then only that provider's key. A failover carries only the fallback's key.
    • The canary tripwire fires. CANARY_TOKEN now goes into the prompt, and it is matched in streamed text too.
    • The raw-text reply fallback runs the response PII scan.
    • GET pass-through survives a compressed reply.
    • An Anthropic caller's own x-api-key is no longer blanked.
  • #196 (merged 2026-10-07, first release): the provider's error reason is relayed to the client, scrubbed.

  • #210 README and docs: LiteLLM 1.106.0 and later ship the built-in guardrail.

Behaviour changes

  • With ALLOW_CLIENT_UPSTREAM_OVERRIDE=true, a virtual-key or OVERRIDE_CLIENT_AUTH request naming any other host now gets a 403. A BYOK caller's own key is unaffected.
  • With ENABLE_CANARY_TRIPWIRE=true, the system prompt now carries CANARY_TOKEN.

pii-leak-benchmark is unchanged at 0.6.0.