Trust

Security

Your agents handle real work: your email, documents, customers and books. Here is where that data goes, who can see it, and how we protect it.

An agent in a plush squid hat locks a steel document cabinet in a small office, while a colleague with a coffee mug looks on.

The short version

  • Your data is stored and backed up in Germany.
  • Conversations, memories and credentials are encrypted with keys of their own.
  • No AI provider may train on your data.
  • One setting keeps all AI processing inside the EU.
  • Your organization's admins see what agents cost, not what they say.
  • Every safeguard is backed up by another.

Where your data lives

The application, database and agent workspaces run on our servers at Hetzner in Germany. Backups are encrypted before they leave the server, stay in Germany, are deleted within 90 days, and are test-restored every week. The master encryption keys live separately, in Google Cloud's key-management service in Frankfurt.

Which AI providers see it

Your agents think with Claude, through Amazon Web Services, and use Google Cloud and OpenAI for images, speech, video and music. None of them may use your data to train their models — a condition of using any provider at all.

For legal and compliance reasons, some providers keep requests for a short time.

Organizations start on “Global”, where processing can run wherever our providers operate. Switch to “European Union” and it all stays in EU regions, for about 10% more; video and music are not available there. One admin setting covers every agent.

Residency covers AI processing, not how messages travel: email goes through our email provider outside the EU, and Teams, Slack and Telegram use their own networks. Web searches and map lookups go to those services. Every company we use is on our sub-processors page.

Encryption

All traffic to Squidler is encrypted. In our database, your conversations, memory, desk content, emails and credentials are encrypted with AES-256, using a data key per agent and per account, wrapped by a master key that never leaves Google Cloud's key-management service.

It protects against a stolen database or backup, or a bug in our access rules: either yields only encrypted data. Workspace files are protected by the sandbox described below and by encrypted backups.

Who can see your data

An agent in a plush squid hat lowers the blind on a glass meeting room, while the customer at the table looks up and smiles.
  • You, and the people you share an agent with, one person at a time.
  • Not your organization's admins. They manage billing and see costs; membership never grants access to an agent's conversations.
  • Squidler staff, only to help you. We look at customer content only to support you or to investigate a problem, and every look is logged before anything is decrypted.
  • Not your agent, for passwords. It refers to a credential by name; the value goes only where you approved, and is masked from what the agent reads back.

Defense in depth

Your data is protected by several independent layers, so a failure in one does not expose it.

In an office corridor, a colleague taps a key card on a reader while an agent in a plush squid hat holds the door.
  1. Every agent has its own sandbox
    An agent's commands run as a separate user with its own memory and processor limits. It cannot see other agents' files or processes.
  2. Agents hold no keys
    The sandbox starts empty: no database passwords, API keys or encryption keys to steal.
  3. Encryption as the backstop
    Even someone who reads the database directly finds only encrypted content.
  4. Connected services get only what you grant
    For Google and Microsoft you choose “Read”, “Write” or “Send” per account, and the provider enforces it.
  5. Only people you allow can instruct your agent
    New Telegram and Teams chats need the owner's approval; Slack guests and outsiders are fenced off; email becomes a conversation only from verified senders you allow.
  6. A careful path to production
    Every change passes automated checks and a dependency audit, and production gets exactly the build already tested on our test environment.

What your agent will and will not do

An agent in a plush squid hat holds out a tablet so the customer can approve something before he goes ahead.
  • It refuses harmful work. The models' own safety systems refuse malware, break-ins and exploits, and we never work around a refusal. Illegal use, fraud and impersonation get an account suspended.
  • It asks first before emailing anyone not on your list, permanently deleting mail, using a tool marked destructive, sending a credential somewhere new, or spending past the daily limit on video and music.
  • It stays within the access you give it. With write access it works without asking each time, so grant what matches your trust; “Read” alone is always an option.
  • It treats what it reads as information. Web page text is marked as information, not instructions.
  • It marks what it makes. Video carries an invisible watermark and content credential, and cannot depict children. Music is marked as AI-generated.
  • It cannot run away. Email, agent-to-agent messages and tool use are capped, and prepaid credits stop a mistake from compounding.

More questions

Need a questionnaire answered, a data-processing agreement or something specific? Write to legal@squidler.io.

Reporting a vulnerability

Found a security problem? Write to security@squidler.io. We read every report and will tell you what we do about it. Our security.txt has the same contact.

See also our Privacy Policy and sub-processors.