Security

Commander is designed for multi-agent workloads that touch code, tools, and untrusted model output. This page is the operator-facing summary. Deep dives: Security Gateway, Sandbox, Multi-Tenancy.

Reporting vulnerabilities

Do not open public GitHub issues for security bugs.

Email sampan090611@gmail.com with:

• Issue type and impact
• Paths / commits involved
• Reproduction steps
• PoC if available

Expect acknowledgment within 48 hours (see product SECURITY.md).

Threat model (summary)

Threat	Mitigations in Commander
Prompt / tool injection	Injection scanning on tool output, sanitizers, reversibility gates for irreversible tools
Secrets leakage	DLP patterns, structured logging without secrets, vault patterns in enterprise gateway
Runaway agents	Approval modes, token budgets, timeouts, circuit breakers
Provider outage / abuse	Failover chains, rate limits, per-tenant quotas
Cross-tenant bleed	Tenant isolation (storage, memory, cache, rate limits)
Supply-chain / deps	CI npm audit; keep keys out of git

Controls you should configure

1. Approval modes

Mode	Use when
plan	Preview only
read-only	Analysis / audit
suggest	Human approves writes
auto-edit	Trusted local dev
full-auto	CI with PR review

export COMMANDER_MODE=read-only

2. API authentication

When running the HTTP server:

export COMMANDER_API_KEY="long-random-secret"

Require Bearer tokens on API routes; do not expose :4000 to the public internet without TLS and auth.

3. Network binding

Default posture favors localhost for the API. In production, put TLS termination and auth at the reverse proxy (Deployment).

4. Local / sensitive code

Prefer Ollama / vLLM so code never leaves your network:

export OLLAMA_BASE_URL=http://localhost:11434

5. Multi-tenant production

Enable tenant provider and quotas — see Multi-Tenancy.

Security benchmarks (in monorepo)

Suite	Purpose
Red Team (47 scenarios)	Attack categories against defenses
AgentDojo	Indirect injection robustness
Tenant isolation fuzz	Cross-tenant mutation tests

Run from the product repo:

pnpm benchmark:redteam
pnpm benchmark:agentdojo

Details: Benchmarks · product BENCHMARK.md.

Architecture map

Request / CLI
  → Auth / rate limit / tenant context
  → Deliberation + topology
  → Agent runtime (tools under policy)
  → Security gateway / sandbox / DLP
  → Quality gates (incl. SAFETY)
  → Result + audit / event log

Checklist before production

• [ ] COMMANDER_API_KEY set; no default example keys
• [ ] Approval mode appropriate for environment
• [ ] Provider keys only in secret store / env
• [ ] Health + metrics endpoints monitored
• [ ] DLQ / compensation reviewed for write-heavy tools
• [ ] Backup strategy for SQLite/Postgres state

Related

• Security Gateway
• Sandbox
• Cookbook: security audit
• Community / responsible disclosure