Built secure by default.

What protects your product data today.

• Shipped

  Content quarantine before any model reads

  Every uploaded document, every customer transcript, every email body, every chat turn is validated and quarantined before a model sees it. A hostile document or a prompt-injection payload in someone's email cannot escalate into your product's state.

• Shipped

  Prompt-injection defense in depth

  Three layers: prompt-level instruction (every user input is wrapped with a defense instruction so the model ignores injected commands), output leakage detection (we scan generated text for leaked system prompts or proprietary methodology and block it), and knowledge leak prevention (only validated items land in your Product Brain).

• Shipped

  Workspace isolation

  Each workspace is per-user. Cross-workspace data access is impossible at the ORM layer: every service function takes a userId and asserts ownership internally. Route-layer auth is defense-in-depth, not the only defense.

• Shipped

  Clerk auth (SOC 2 Type II provider)

  Authentication is handled by Clerk, a SOC 2 Type II-certified identity provider. Session tokens are short-lived. Password reset and 2FA are standard.

• Shipped

  Encrypted at rest, in transit

  Postgres at rest is encrypted. Your Product Brain storage is encrypted. All traffic is TLS. Document uploads pass through size and MIME-type validation before any extraction runs.

• Shipped

  We do not train on your data

  Your product evidence, your customer signals, your decisions stay in your workspace. They are never used to train models we ship to other customers. The methodology improves; your data doesn't leave.

• Shipped

  OAuth-first agentic surface

  The agentic surface (native MCP server + plob CLI) ships the way it should have been built across the category: OAuth-first authentication and short-lived API keys, no static shared secrets, read-only by default, with each write scope an explicit opt-in.

• Shipped

  Scoped permissions per workspace

  A Brain command call from an external agent sees only the workspace its connection is bound to. The binding and scope set are re-validated on every call, and revocation takes effect on the next call. No cross-workspace leakage. No 'one token reads everything.'

• Shipped

  Audit log per Brain command

  Every Brain command call (actions, reads, and writes — including denied attempts) is recorded with timestamp, agent identity, and workspace. You can audit who asked what, when, from where.

OWASP ASI coverage.

The OWASP Agentic Security Initiative defines the threat surface for LLM-enabled software. ProductLobster maps to the top items in production today.

• ASI01: Prompt injection

  Defense in depth: input wrapping with defense instruction, output leakage detection, Product Brain validation gate.

• ASI02: Insecure output handling

  Output is structured, validated, and rendered with sandbox restrictions on prototype output (iframe sandbox without same-origin).

• ASI06: System prompt leakage

  Three-tier risk scoring on Product Brain deposits; pattern-match detection on generated output before it reaches the user.

• ASI08: Vector and embedding weaknesses

  Circuit breakers on retrieval; content validation gate before any Product Brain deposit; never-throw service contract isolates upstream failures.