Security & Risk Boundaries
Security through bounded authority and explicit gates.
Autonomous systems fail catastrophically when granted open-ended authority. IRISH enforces defense-in-depth: specialists are bounded by construction, prompt inputs never grant extra privileges, and code publication is locked behind human owner approval.
Defense-In-Depth
Core Security Guarantees
Local Boundary Enforcement
The IRISH supervisor runs locally. Source files are read directly from local git repositories and never sent to hosted database clusters or shared multi-tenant SaaS backends.
Prompt Injection Mitigation
Instructions provided in issues or chat cannot elevate privileges, override file scope locks, or trigger remote actions. Instructions are parsed as data within the bounded specialist schema.
Zero Silent Remote Writes
Direct pushes to primary git branches (main/master) are blocked at the daemon layer. All code publication takes the form of reviewable GitHub pull requests following explicit owner approval.
Secret Sanitization
Diff outputs and intermediate logs are scanned by automated heuristic analyzers to detect and prevent leakage of API keys, tokens, environment secrets, or private credentials.
Network Isolation
Isolation of Privileged Control Surfaces
DCW maintains a strict architectural policy: privileged control endpoints such as mcp.dcw.co.in (Model Context Protocol endpoint) and auth.dcw.co.in (Owner OAuth server) are private control mechanisms that require authenticated, signed access.
These privileged hosts are completely isolated from this public website (irish.dcw.co.in) and dcw.co.in. They share no credentials, sessions, or execution paths.
Vulnerability Reporting
Coordinated Disclosure
If you discover a security vulnerability in any DCW or IRISH software, report it directly through our contact channel. We appreciate responsible disclosure and investigate all valid reports promptly.