// legal · safety · security

Legal notices, safety & security

This page states legal disclaimers, project independence (including explicit non-affiliation with Google, Google DeepMind, and related entities), and security expectations for Bumblebee — community-developed open-source software — and for this website. It is meant for operators, deployers, integrators, and visitors evaluating the project. Unless you separately enter into a written agreement, nothing here creates a binding contract; it is a prudent-use notice for experimental agent software.

Independence, community & third parties

How Bumblebee fits in your stack

Typical setups keep inference on your hardware (for example via Ollama) and store configuration, memory, and logs on disk under your user. Optional features may call third-party APIs (search, crawl, cloud inference) when you explicitly configure keys and tools — those requests are subject to the provider’s terms and privacy policy, not ours.

The project is open source so you can inspect behavior, pin versions, and fork. That does not remove operational risk: you are responsible for how you deploy, network-expose, or share access to a running entity.

Known safety & security considerations

An honest list beats pretending the risks do not exist. Mitigations are suggestions, not guarantees.

  • Prompt injection & untrusted content Email, web pages, attachments, or chat text can contain instructions meant to steer the model. If the agent can run tools or send messages, that content may indirectly cause actions. Prefer allowlists for chat surfaces, constrain tools, and avoid piping high-value secrets through untrusted prompts.
  • Tools, MCP servers & automations Skills, scripts, MCP integrations, and automations run with whatever OS and account permissions you grant the process. Only enable code and servers you trust; review updates before upgrading in production-adjacent environments.
  • Messaging channels (Telegram, Discord, CLI, …) Anyone who can message your bot may be able to trigger tool use depending on your configuration. Use private bots, user allowlists, and network controls so the harness is not reachable by the whole internet unless that is intentional.
  • Secrets & credentials Tokens and API keys in .env or config files are only as safe as the host machine and backup policy. Use scoped keys, rotation, spend limits where providers allow them, and separate accounts for experiments vs production.
  • Hybrid / remote workers If you split “brain” and “worker” across networks, you introduce trust boundaries (tunnels, queues, object storage). Compromise of any hop can affect the whole pipeline — design authentication, TLS, and firewall rules accordingly.
  • Model behavior & outputs Local models can still produce incorrect, biased, or harmful text. Do not rely on the harness for safety-critical decisions without human oversight and domain-specific validation.

Responsible use

Do not use Bumblebee to break law, harass people, scrape or send data you are not entitled to, or bypass security controls. You are responsible for compliance with privacy, employment, sector, and export rules that apply to your jurisdiction and data. This software does not provide legal, medical, financial, or therapeutic advice; outputs are not a substitute for qualified professionals.

Reporting security vulnerabilities

If you believe you have found a security vulnerability in Bumblebee, please report it responsibly before public disclosure. Prefer the repository’s GitHub Security Advisories flow (private report) so maintainers can reproduce and ship a fix. Include enough detail to reproduce, affected versions, impact assessment, and — where safe — a minimal proof of concept.

Please do not access, modify, or exfiltrate user data without explicit authorization; do not perform testing that could degrade third-party services. Coordinated disclosure helps protect downstream operators who run this software in production.

This project is a volunteer-driven open-source effort: there is no bug bounty or formal SLA by default. We still ask that you allow a reasonable window to address valid reports before publishing exploit details.

FAQ

Is Bumblebee affiliated with Google, Google DeepMind, or Alphabet?

No. Bumblebee is community-maintained open-source software. It is not a product, service, or official project of Google, Google DeepMind, Alphabet, or their affiliates. Mentions of Gemma or related names describe optional third-party models you may run yourself; they do not imply endorsement or responsibility by those companies. See Independence, community & third parties above.

Does Bumblebee send my chats to a vendor by default?

With a purely local inference setup, prompts and completions stay on your machine (subject to how you configure Ollama or other runtimes). If you point the harness at cloud APIs or enable tools that call the network, those components may transmit data according to their policies.

Is this software audited or certified?

There is no guarantee of third-party security audit or certification (e.g. SOC 2). Self-hosted software puts the operating environment under your control — and your responsibility.

Can I restrict who talks to my bot?

You should configure platform-specific allowlists, private bots, or network rules so only intended user IDs or networks can reach the presence layer. Treat a reachable bot as remote code execution surface if tools are enabled.

What if I uninstall?

Removing the package does not necessarily delete configs, memory stores, or logs on disk. Follow project documentation to wipe sensitive directories and revoke tokens when decommissioning.

← back to home