Open-source design agent: on-brand stills, decks, diagrams, charts and motion graphics from a chat thread, built on Deep Agents.
Security policy
Reporting a vulnerability
Use GitHub's private vulnerability reporting on this repository ("Report a vulnerability" under the Security tab). Do not open a public issue for a security problem, and do not include tokens, thread contents or files people shared in a report.
What counts
- Any path that lets model output, a Slack message, a file someone shares, Paper canvas content, MCP metadata or a web page change policy, reveal a secret, reach a tool outside the assembled surface, or send anything outside the thread.
- Credentials reaching the sandbox, a prompt, graph state, a log, an error, a Slack block, a fixture or a snapshot.
- Sandbox escapes: reading or writing outside
/workspace, reaching the Docker socket, the host, cloud metadata, or the network when egress is denied. - Server-side request forgery through
fetch_assetor any URL the agent is given. list_tracesorread_tracereading outside the run's own LangSmith project, or writing to LangSmith at all.- Paper scope bypass: writing to the protected brand file, deleting nodes or writing tokens when
disabled,
paper-gen:/paper-asset:/file:URLs reaching the canvas. - Delivering a file twice, delivering a file the thread did not produce, or claiming a delivery that did not happen.
- Bridge or API authentication bypass, DNS rebinding against the Paper bridge, thread ownership bypass on the self-hosted API, or Slack signature verification bypass.
Trace access
The deployed agent can read every thread's traces in its own LangSmith project (list_traces,
read_trace), with the host's credentials, so it can find why a job failed or what a helper did. A
trace holds what people asked, what they shared and what the agent replied. Anyone who can message
the agent can therefore ask it about other people's requests, DMs included. Where that is not
acceptable, set DESIGN_AGENT_ENABLE_TRACES=0. The tools only read, only from the project the run is
traced to, and hand trace contents to the model as data, never as instructions.
What we commit to
- Acknowledge reports within five business days.
- Keep defaults safe: sandbox egress denied, Paper off, deletes and token writes off, nothing published or sent outside the thread.
- Publish fixes with a changelog entry and, where relevant, a regression test in
tests/.
Supported versions
Only the main branch is supported before the first tagged release.