Prompt-injection & content-sanitization defense

[JSONbored]

Problem: A customer's own repo content (issue text, code comments, README) becomes untrusted input the moment it reaches an autonomous agent acting with real authority — today nothing sanitizes or hardens against adversarial content in that path.

Area: Platform / Security

Proposal: Add a content-sanitization layer between anything read from a target repository and the agent's own instructions/context, and add defenses against content that attempts to redirect the agent's behavior.

Deliverables:

• A sanitization/defense layer applied to all repo-sourced content before it reaches the agent.

Acceptance criteria:

• A test repo containing adversarial content designed to redirect agent behavior does not succeed in doing so.

Test scenarios:

• Run the loop against a test repo deliberately seeded with injection-style content in an issue and in a code comment.

Boundaries:

• Blocked on Generalized plan → implement → test → submit loop #4786.

---

Part of #4778.

[linear-code]

JSO-3467