Skip to content

Security: VoyagerEddy/AutoResearch

Security

docs/SECURITY.md

Security boundaries

AutoResearch handles secrets, accesses networks, generates code, and can execute remote commands. It applies these boundaries by default:

  • .env, account data, datasets, paper PDFs, and experiment results are excluded from Git.
  • The browser settings API never echoes secrets, and SSH passwords are never stored in the database.
  • Generated files are confined to the project workspace. Absolute paths, traversal, common sensitive filenames, conflicts, and oversized output are rejected before any file is written.
  • Automatic PDF downloads allow only HTTPS arXiv hosts and at most 25 MB per file.
  • Remote uploads exclude .git, .env, datasets, results, and virtual environments.
  • AutoDL provisioning requires billable-action confirmation. Releasing a replacement instance requires separate authorization.
  • GPU recovery attempts are bounded and cannot continuously provision billable instances.
  • Git remotes accept only standard GitHub HTTPS or SSH URLs and reject embedded tokens.
  • MCP listens on 127.0.0.1 by default. ChatGPT connects through outbound Secure MCP Tunnel without opening an inbound port.
  • MCP tools never return OpenRouter keys, AutoDL tokens, or SSH passwords. AutoDL SSH data is converted directly into an in-process connection object.
  • Billable provisioning and remote execution require confirm_billable and confirm_execute, respectively. Tool descriptions require explicit consent in the current conversation before those fields are set.

This is not a strongly isolated sandbox. Starting a remote experiment authorizes the command displayed in the UI. Review generated/ before execution, use disposable AutoDL instances, and back up important data.

Rotate any secret that has appeared in a chat, screenshot, or plaintext requirements document. Store the replacement only in local settings.

There aren't any published security advisories