Kronos Agent OS (KAOS) is a self-hosted agent runtime. It can connect to LLMs, local memory, MCP tools, scheduled jobs, browser tools, Telegram/Discord, and optional infrastructure tooling. Treat it like a local automation system with access to whatever credentials and tools you enable.
Security fixes target the latest public release and main until the project has a broader release policy.
Do not open public issues containing secrets, exploit details against real systems, or private infrastructure data. Report security-sensitive issues privately to the maintainer first.
Include:
- Affected version or commit.
- Configuration used, especially enabled capability gates.
- Minimal reproduction steps.
- Impact and expected behavior.
Fresh installs are conservative:
ENABLE_DYNAMIC_TOOLS=falseREQUIRE_DYNAMIC_TOOL_SANDBOX=trueENABLE_MCP_GATEWAY_MANAGEMENT=falseENABLE_DYNAMIC_MCP_SERVERS=falseENABLE_SERVER_OPS=false
These defaults are intentional. Enable risky capabilities only in trusted local environments.
When enabling dynamic tools, build the local sandbox image with scripts/build-sandbox.sh first. kaos doctor reports a hard failure if ENABLE_DYNAMIC_TOOLS=true but Docker or kronos-sandbox:latest is unavailable.
Primary risks:
- Prompt injection through user messages, web pages, email, Telegram, Discord, and external documents.
- Secret leakage from
.env, MCP server env vars, logs, tool arguments, and model output. - Unsafe dynamic code or tool execution.
- Unsafe runtime mutation through dynamic MCP server registration.
- Accidental infrastructure actions through server ops tools.
- Scheduled jobs repeating a harmful or expensive action.
- Dashboard exposure beyond localhost without proper auth and network controls.
Security controls currently include input validation, external-content sanitization, output redaction, PII masking for logs/traces/audit previews, loop detection, cost guarding, capability gates, and sandbox-required dynamic tool execution.
See docs/SECURITY.md for implementation details.
KAOS is intended for user-authorized personal, research, productivity, and operations workflows. Do not use it to access systems without permission, bypass security controls, harvest credentials, or run automation against third-party services in ways that violate their rules.