Security & Safety¶
Simple Explanation¶
Run many small steps safely: secrets never leak in logs, dangerous code is blocked, and we enforce time and resource limits.
Technical Explanation¶
Security combines template sanitization, secrets/PII redaction, sandboxed code execution (restricted builtins, timeouts, memory/FD limits, workspace path filters), and operational guardrails (cost/concurrency/rate limits, idempotency). See the concise threat model below.
Built-in protections¶
- Secrets redaction: prevents credentials in logs (installed automatically)
- PII filtering: sanitize templates to avoid leakage
- Template sanitization: block dangerous tokens/patterns
- Sandboxed execution: restricted builtins/modules, timeouts
Threat model (concise)¶
- Trust: code/config from trusted operators; LLM providers trusted; user inputs untrusted; external APIs/vector/cache stores treated as potentially hostile.
- Code execution risks: mitigated by sandbox (no network, restricted builtins, timeouts, memory/FD limits, workspace path filter).
- Resource exhaustion: timeouts, memory/FD limits, rate limits, quotas on workspace/files.
- Information disclosure: env filtering, workspace isolation, redaction in logs/outputs.
- Privilege/supply chain: role checks on sensitive ops, pinned deps; prefer managed secrets (Vault/Key Vault) and signed skills where available.
Operational checklist - Use Managed Identity/Key Vault (avoid plain API keys in prod). - Enable metrics/logging with redaction; alert on error/timeout spikes. - Set cost/concurrency/rate limits for fan-out workloads. - Rotate secrets regularly; restrict network egress for sandboxes where possible.
Operational guardrails¶
- Cost caps and concurrency limits
- Failure-rate thresholds and total duration limits
- Idempotency cache: avoid duplicate execution
See: - Logging Helpers - Sandbox Execution - Parallel Agents