IncidentAgent behavior
DeepSeek Harness: a sandboxed agent could switch off its own sandbox through an unauthenticated local API
DeepSeek Harness is an open-source, local-first tool for running AI coding agents, released in August 2026 with a browser interface backed by a local HTTP API on port 3080. OX Research (Nir Zadok and Moshe Siman Tov Bustan) found that the API's authentication check looked only at the HTTP Host header and never at where the connection came from. The operating system sandbox the agent ran in allowed loopback networking, so a confined agent could call the API with one shell command and raise its own session to the "danger-full-access" mode, which disables the sandbox. The flaw was confirmed on August 24, fixed in 0.1.2-alpha.1 on August 27, verified by OX on August 30 and published as CVE-2026-82533 on September 8, 2026.
Replay the attack step by step in the Learning Center
How it happened
- The agent runs inside an OS sandbox that blocks file writes outside the project but allows connections to localhost.
- The local control API accepts any request whose Host header looks local, without checking the connection's peer address.
- One curl command from inside the sandbox changes the session's permission mode to danger-full-access, and the sandbox is gone.
Impact
Any prompt injection or misbehaving agent in an affected version could escape confinement and act on the host with the user's rights; no exploitation in the wild was reported.
The fix
Upgrade DeepSeek Harness to 0.1.2-alpha.1 or later, including copies bundled inside desktop wrappers, and keep local control APIs behind real authentication.
Lessons
- A sandbox that can be reconfigured from inside it is not a sandbox; the control plane must be out of the agent's reach.
- A Host header is not an identity: local services that make security decisions need an authenticated caller, not a lookalike address.