What works today, and what comes next

Dates are targets, not commitments. The Security status page records what is tested at each stage.

Available today

• Per-agent, per-engagement credentials, issued and revocable by the firm
• Allow / Deny / Elevate enforcement, with named-human approval for sensitive actions
• Hash-chained, tamper-evident audit trail on the firm's infrastructure
• Works with OpenAI, Anthropic, Google, and self-hosted Llama and Mistral
• Single-device deployment for pilots of up to a few dozen agents; open-source codebase

Target: Q1 2027

• Remote MCP gateway: agent platforms add MahaGuardian as an MCP integration, with tools limited to one engagement
• OAuth 2.1 authorisation for platform clients, issuing short-lived engagement-scoped tokens
• Approval flow for named approvers, separate from the platform's own confirmation step
• Connectors for Microsoft 365, Google Drive, Slack and Confluence alongside iManage and NetDocuments
• Published end-to-end demonstration

Target: H1 2027

• Always-on, firm-hosted Guardian
• High-availability key management, backed by HSM or cloud KMS, with the keys still held by the firm
• One Guardian per customer tenant, so a platform can serve many firms without holding their keys
• Audit export to SIEM tools
• Independent penetration test and security review

Target: H2 2027

• Federation between Guardians, so agents on different sides of a transaction can work in a shared space without either side's keys leaving its custody
• Certification path — ISO 27001 or SOC 2
• EU-hosted managed option for firms that prefer not to operate the Guardian themselves