Skip to content

Security: B-A-M-N/Conflux

Security

SECURITY.md

Security

Security Model

Conflux follows a constrained-orchestration model:

  • Treat generated implementations as untrusted until gates pass.
  • Centralize write governance through contracts and mutation checks.
  • Route external MCP usage through one broker with policy + backend profiles.
  • Keep observability transport separated from mutation permissions.

Key Controls

  • Path isolation: safe path validation under project root.
  • Track contracts: write-scope and quality-gate constraints.
  • Sentinel gates: structural validation + adversarial checks.
  • Red/Green + Smoke gates: enforce project-level regression barriers.
  • Forensics lane: reproducible failure history and fingerprinting.
  • Governance provenance: policy version/hash + metrics hash logged per admission decision.
  • Control-loop rollback: automatic parallelism rollback on post-scale health regression.
  • External MCP policy: explicit backend/tool allowlists.
  • Backend registry: per-backend egress/required-env/preflight/circuit-breaker config.

External MCP Hardening

Use both layers:

  1. Runtime backend profile (set_external_mcp_backend)
  2. Tool authorization policy (set_external_mcp_policy)
  3. Start from reproducible templates:
    • conductor/templates/mcp_backends.example.json
    • conductor/templates/external_mcp_policy.example.json

Recommended defaults:

  • Use narrow egress_allowlist per backend.
  • Set required_env for tokenized providers.
  • Run preflight_external_mcp_backend before enabling production calls.
  • Keep strict mode optional:
    • CONFLUX_EXTERNAL_MCP_STRICT_MODE=0 (default lower-friction)
    • CONFLUX_EXTERNAL_MCP_STRICT_MODE=1 (high-assurance mode)

Optional MCP availability behavior:

  • CONFLUX_OPTIONAL_EXTERNAL_MCP_SOFT_FAIL=1
  • CONFLUX_OPTIONAL_EXTERNAL_MCP_BACKENDS=code7,context7

This allows workflows to proceed when optional augmentation is unavailable.

Vigilance Alignment (Recommended)

  • Bind Vigilance locally by default (127.0.0.1).
  • Require auth/signature on ingest/control channels.
  • Keep dashboard/SSE read-only; isolate control endpoints.
  • Apply rate limits and bounded queues.
  • Use a trusted relay path if worker sandboxes are strict.

There aren’t any published security advisories