Smarter runs LLM applications that hold API keys, customer data and cloud credentials, so we treat security reports as our highest priority. Thank you for helping keep Smarter and its users safe.

Please do not report security vulnerabilities in public GitHub issues, discussions or pull requests.

Report them privately, in either of these ways:

1. GitHub private vulnerability reporting (preferred). Open the repository's Security tab and choose Report a vulnerability. The report, and our discussion of it, stay private until an advisory is published.
2. Email lpm0073@gmail.com, with [smarter security] in the subject.

Please include as much of the following as you can:

• the affected component, e.g. the REST API, the web console, a manifest kind, the Helm chart or the Docker image, and the affected versions,
• the impact: what an attacker can do, and with what access,
• steps to reproduce, a proof of concept, or the request and response,
• any known mitigation.

We will keep you informed of our progress, and may ask you for more information. When the fix is released, we publish a GitHub Security Advisory, request a CVE when one is warranted, and credit you, unless you prefer to remain anonymous.

We ask that you give us a reasonable time to release a fix before you disclose the vulnerability publicly. If we cannot fix it within 90 days, we will agree on a disclosure date with you.

Security fixes are released as patch releases of the latest minor version, and are announced in the changelog and the advisory. The Docker image on Docker Hub and the Helm chart are published with every release.

In scope:

• this repository: the Django platform, its REST API, the web console and its React apps, the manifest brokers, and the Celery tasks,
• the Docker image and the Helm chart that this repository publishes,
• the default configuration that they ship with.

Out of scope:

• vulnerabilities of third-party dependencies that Smarter does not expose; please report them upstream, and tell us if Smarter is affected,
• the behavior of LLM providers, and prompt injection that does no more than the prompt's author could do anyway; prompt injection that crosses an account boundary, leaks a secret, or calls a tool the user may not use is in scope,
• findings that need a compromised administrator account, physical access, or a misconfiguration that the documentation warns against,
• denial of service by volume, missing security headers without a demonstrated impact, and reports from automated scanners without a proof of concept,
• the hosted services at smarter.sh, other than through a vulnerability of this repository's code.

We will not take legal action against, or ask law enforcement to investigate, anyone who researches and reports a vulnerability in good faith under this policy. Good faith means that you:

• test only against your own installation, or accounts you own,
• don't access, change or delete other users' data, and stop and report as soon as you encounter any,
• don't degrade the service for others, e.g. by load testing or spam,
• give us a reasonable time to fix the issue before you disclose it.

Operators of a Smarter installation should also read the security documentation. In particular:

• set SMARTER_SECRET_KEY to a long, random value that every pod shares; outside of the local environment, Smarter refuses to start without one,
• keep the API rate limits, SMARTER_THROTTLE_RATE_ANON and SMARTER_THROTTLE_RATE_USER, enabled,
• keep secrets out of manifests and source control: store them as Smarter Secrets, or in your cloud's secret manager,
• upgrade to each patch release of your minor version promptly.