Advisories for Pypi/Atomic-Agents-Stack package

_estimate_batch_cost (atomic_agents/agent.py) looks up the per-model output price with PRICING.get(model, {}), returning 0.0 for any model not in the hardcoded pricing table. _check_batch_reservation then early-returns when the reservation is <= 0, skipping the batch reservation entirely. That reservation is the only defense against the documented fan-out race where every parallel helper/delegate reads the identical pre-batch on-disk cost total and each passes its individual check even though the collective spend overruns …

The HTTP MCP server-registry backend factory (atomic_agents/mcp_registry/http.py, make_http_mcp_server_registry_backend_from_url) accepts both http and https schemes. Catalog entries carry command/args that are type-validated but content-unrestricted, and are later spawned as local stdio subprocesses by MCPClientPool. Over a cleartext http:// catalog URL, a network man-in-the-middle can rewrite the catalog response to inject an arbitrary command/args and obtain code execution on the agent host, with no LLM involvement. The Policy MCP allowlist is not …

The optional dashboard HTTP server (atomic_agents/dashboard/serve.py) builds filesystem paths directly from the request path and serves them without a containment check. It is the only per-request untrusted-path site in the codebase that does not route through _io.safe_resolve_under. Literal ../ segments survive urlparse and Path joining, so a request can read files outside the intended agents_root (including via the static branch). Impact: arbitrary file read. The default bind is loopback, but …