# Security

## Prompt and repo safety

`fable-agent` may read untrusted markdown, tickets, repos, and model output. Treat all of it as data, not instructions.

- Do not paste, vendor, replay, or execute leaked proprietary system prompts.
- Do not use permission-bypass flags.
- Send model prompts through stdin or structured APIs, not shell-interpolated command strings.
- Keep secrets in environment variables; never hardcode fallback API keys or local key paths.
- Run code-changing agents in scoped workspaces/worktrees when possible.
- Keep docs and CLI parity green with `npm run -s docs:check` before publishing.

## Defensive Fable-style profile

Allowed: multi-model fan-out plus judge synthesis, citations for factual/current claims, refusal/reformulation for harmful requests, bounded loops with validation evidence.

Not allowed: leaked prompt reproduction, prompt-extraction recipes, jailbreak/sandbox-bypass instructions, autonomous execution without permission boundaries.

## Reporting

Open an issue with reproduction steps, affected version/commit, and impact. Do not include live secrets or proprietary prompt text.