Debo is a private AI memory and context operating system designed around uncompromising user privacy, verifiable source grounding, and deterministic access controls. Because Debo handles deeply personal memories, documents, and credentials, security is a core pillar of our engineering discipline.

If you discover a security vulnerability within Debo, please do not open a public GitHub issue. Instead, report it directly to our security team:

• Email: security@debo.life
• Response Target: Initial triage within 24 hours; formal status update within 72 hours.
• Coordinated Disclosure: We adhere to a 90-day coordinated disclosure timeline, providing fixes prior to public disclosure.

When reporting, please provide:

1. A description of the vulnerability and its potential impact.
2. Step-by-step reproduction steps or a minimal proof-of-concept (PoC).
3. The affected component (debo-web, debo-mcp, debo-worker, debo-document-worker, or package).
4. Any potential mitigations you have identified.

The Large Language Model is an untrusted reasoning component and must never be the authority for permissions. All access checks (multi-tenancy boundaries, role-based access, tool authorization, source retrieval, and destructive action approval) are performed by deterministic application code in Debo before any query or tool execution occurs.

All persistent records in Neon PostgreSQL, Cloudflare R2, and vector/memory stores are strictly partitioned by workspace_id. Cross-workspace data leakage is treated as a P0 critical security defect.

All ingested content (web pages, PDFs, emails, voice transcripts, third-party connector events) is treated as potentially adversarial. Ingested text is sanitized and wrapped in strict data delimiters (<source_data>...</source_data>) to mitigate indirect prompt injection attacks before being exposed to model contexts.

Agents cannot independently execute destructive operations (permanent deletion, bulk data export, modifying external systems via connectors, public sharing) without explicit, auditable user approval recorded in the Approval log.

• Third-party OAuth tokens and API keys are stored with envelope encryption and never logged or exposed to the frontend.
• API keys for Debo MCP access use cryptographically secure random tokens hashed with SHA-256 before database storage.
• No production secrets or .env files are permitted in version control.

For detailed analysis of our threat vectors, mitigation strategies, and zero-trust data flows, refer to:

• docs/security/README.md
• docs/security/threat-model.md
• docs/security/secrets.md
• docs/mcp/security.md