Skip to content

Secret redaction and prompt-injection defences #30

Open
Open
Secret redaction and prompt-injection defences#30
Labels
P3laterenhancementNew feature or requestphase-3AgentMux integration

Description

@agent3-workflow
agent3-workflow
bot
opened on Sep 26, 2026

Do. Redact known secret shapes from tool output before it reaches a cloud model; treat web and tool output as untrusted (tag it; don't follow instructions inside it). Optional: a reviewer model for risky tool calls (Goose "adversary", Codex "Guardian"). OS sandboxing is deferred; AgentMux container runs cover isolation meanwhile.

Spec §6.7.

Spec: SPEC_MUXCODE_HARNESS_AND_FEATURE_GAPS_2026_09_26.md

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    P3laterenhancementNew feature or requestphase-3AgentMux integration

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions