How CyberPocket handles AI, your data, and your trust signals

CyberPocket sends your alert text (and optional knowledge-base excerpts you have explicitly added to your workspace) to a hosted large language model to produce a structured analysis: risk level, evidence extracted, plain-English explanation, and next steps.

• Every output is advisory. A human analyst must verify findings against original telemetry before acting.
• The model is constrained to defensive cybersecurity scope only. It refuses requests to generate exploits, malware, or attack instructions.
• User-supplied text is wrapped as untrusted data in the prompt — attempts to "ignore previous instructions" are treated as content, not commands.
• We do not use your prompts to train third-party models.

Your input is sent over TLS to our server, where a redaction pass strips tokens, keys, JWTs, credential-shaped strings, SSN/card-number patterns, and other obvious secrets before any AI model sees it. The exact rules live insrc/lib/redact.tsand run on every analyzer call.

Above every alert input you'll see a yellow banner that asks you not to paste passwords, customer PII/PHI, or production credentials. This warning exists because:

• The redactor catches common patterns but cannot guarantee 100% coverage on arbitrary free-form text.
• Your saved analyses live in our database under row-level security; they are still safer when they do not contain raw secrets in the first place.
• If you ignored the warning, you can delete the affected analysis from your history at any time.

• Database: hosted Postgres, encrypted at rest. Every user-owned table has Row Level Security; you can only read rows you own (or that belong to your workspace).
• Uploaded files: stored in a private bucket. Public links are never generated; downloads happen through signed URLs scoped to your session.
• Logs: server logs are passed through a redaction helper (src/lib/log-redact.server.ts) and truncated to bounded length so raw alert content is not retained in plain text.

The app records a privileged-action trail in public.audit_logs when an admin or owner:

• Deletes an individual analysis or clears all history
• Marks user feedback on AI output as reviewed
• Removes a knowledge-base entry

Admins can browse pending AI-output flags from the internal review queue. That queue is the source of truth for "needs human review" findings; approvals and dismissals are written to the audit log with the actor's user id and timestamp.

CyberPocket provides redaction, structured outputs, refusal of offensive prompts, row-level security, audit logging, and per-row delete. You remain responsible for: not pasting production secrets, reviewing AI output before acting, granting admin roles only to people you trust, and complying with your organization's data-handling policies.

We do not currently hold SOC 2, ISO 27001, or HIPAA attestations. We document our controls openly so enterprise buyers can evaluate them honestly. The notes below describe the controls we operate today and the formal attestations we are preparing for — they are commitments and roadmap markers, not certifications.

• SOC 2 Type II readiness (in progress) — control mapping complete across Security, Availability, and Confidentiality trust criteria; auditor selection underway; observation window targeted to begin once policies are signed off. Placeholder for the Type II report; pre-audit gap assessment available to qualified prospects under NDA.
• ISO 27001 readiness (planned, Year 2) — ISMS scope drafted around the analyzer service and supporting infrastructure; Statement of Applicability and risk register are maintained internally and updated quarterly.
• Audit logs — privileged actions (admin role changes, data exports, deletions, redaction-rule edits) are recorded server-side in public.audit_logs with actor, target, timestamp, and request metadata. Logs are append-only from the application path and retained for 12 months.
• Incident response — on-call rotation with a 1-hour acknowledgment target for Sev-1, customer notification within 72 hours of confirmed material incidents, and a written post-mortem published to affected workspaces. The current runbook and status page are linked from the in-app help menu.
• Data Processing Addendum — a standard DPA is available on request for paid workspaces; EU data residency and custom retention windows are available on the Business tier.

If you need a formal attestation, signed DPA, or a security questionnaire response before adopting CyberPocket, email trust@cyberpocket.org and we'll tell you straight what we can and cannot sign today.