Forge is a prompt-engineering tool for red-teamers, alignment researchers, and model providers who need to author, refine, and evaluate persona-layer system prompts. It exists so defenders can build and stress-test their own prompts — and the prompts of models they are authorized to evaluate — before adversaries do it for them.

Use Forge only to build prompts for:

• models and endpoints you own or operate, or
• targets you have explicit written authorization to test (a provider's red-team program, a bug-bounty scope, an internal safety evaluation).

Do not use Forge to attack third-party services without permission, to generate operational content whose only value is real-world harm, or in a way that violates the target provider's terms of service or applicable law. You are responsible for how you use it.

Forge generates persona-layer system prompts and stores drafts, learned context, and API keys under ~/.forge/ on your machine. Nothing there is committed by this repo — the .gitignore blocks .env, keys/, *.key, config.json, memory.json, and saved/ from ever entering the tree by accident. On macOS and Linux the keys directory and key files are chmod'd to owner-only (0700 / 0600); on Windows the same call is a no-op and the file inherits the profile's NTFS ACL.

Treat any prompt you generate as sensitive material. Handle drafts the way you would handle any other red-team artifact: keep them off shared drives, out of public repos, and away from chat services that log prompts.

If you use Forge to refine or emulate a target's leaked system prompt, treat that extracted prompt as the provider's confidential material — report it through their responsible-disclosure channel and do not publish it. Forge's reference and emulation features are for authoring against a target's dialect, not for redistributing the target's IP.

If you find a security issue in Forge (a secret-leak path, an unsafe default, a TLS-verification bypass, a Textual-render crash reachable from model output, etc.), please open a private report rather than a public issue: use GitHub's "Report a vulnerability" (Security Advisories) on this repository. Include repro steps, affected version (__version__ in forge_core.py), platform, and impact.

We aim to acknowledge within 72 hours and ship a fix or documented mitigation on the next release cut. Do not open public issues or PRs describing the vulnerability until a fix has landed.

Only the latest tagged release (see pyproject.toml version) is supported for security fixes. Pin from a tag if you need reproducibility.