GitHub reviewed GHSA-pfwx-xjp4-qxcj / CVE-2026-69157 on November 18, 2026: atomic-server exposes a well-known /setup invite that grants unauthenticated clients write access to the entire Drive on a fresh instance. The reporter showed a GET with an attacker public-key creates a new Agent and adds it to the Drive write array—CVSS 3.1 9.8, no patch release listed yet for versions 0.29.0 through 0.40.1.
Teams that let coding agents spin up Docker Compose stacks for personal knowledge bases or MCP-adjacent data stores should treat this like any other exposed admin bootstrap URL.
TL;DR
| Question | Answer |
|---|---|
| Am I affected? | Yes if you run atomic-server ≤0.40.1 with /setup reachable from untrusted networks |
| Is there a fix? | No released patched version in the advisory as of Nov 18, 2026—use network controls |
| Auth required? | No for invite redemption |
| Fresh install risk? | Full Drive takeover before a legitimate admin claims the server |
| Agent angle? | Compose/IaC files agents write may publish the port by default |
Mechanism step by step
From the GitHub advisory and reporter luuhung1217:
server/src/appstate.rsprovisions a well-known invite at{server}/setupwithwrite = true,target= Drive root,usagesLeft = 1.server/src/plugins/invite.rsredeems invites withfor_agent: _—the redeemer identity is discarded, so possession of the URL + key material is sufficient.- Attacker calls
GET /setup?public-key=<attacker_ed25519>withAccept: application/ad+json(no auth headers). - Server returns a Redirect confirming the new Agent URL.
GET /on the Drive shows the attacker Agent in thewritearray alongside legitimate root writers.- Attacker signs a normal Commit (ed25519 over JCS-canonical JSON) and
POST /commit—Drive root edit succeeds.
Expected behavior: /setup should require operator authentication or a one-time secret printed at bootstrap.
Actual behavior: Any unauthenticated network client can redeem and gain Drive-wide write.
The advisory also documents related flaws for a separate chain: with write access, attackers may read server root private keys and impersonate root (reported separately), plus TOCTOU on usagesLeft and an inverted invite-expiry check in invite.rs. Patch planning should assume defense in depth, not “one line fix.”
Who is affected
| Field | Detail |
|---|---|
| Package | atomic-server (Rust), ontola/atomic-server |
| Affected | ≥ 0.29.0, ≤ 0.40.1 |
| Patched | None listed (Nov 18, 2026) |
| CWE | CWE-306 — missing authentication for critical functionality |
| Reporter proof | joepmeneer/atomic-server:develop Docker on localhost:9883, fresh store |
Self-hosters using atomic-server as a linked-data backend, personal wiki, or research datastore—including stacks a Claude Code or Codex session generates—match the threat model if port 9883 (or your mapped port) is on the LAN or internet.
Remediation and checks
Vendor remediation text in the advisory:
- Require a one-time secret (stdout at startup) or operator authentication before
/setupredemption. - Invalidate
/setupafter first use with an atomicusagesLeftdecrement—current read-check-decrement is TOCTOU. - Fix the inverted expiry check (
if expires > now { reject }is backwards per reporter).
Until a release ships:
docker inspect <container> --format '{{json .HostConfig.PortBindings}}'
Confirm atomic-server is not bound to 0.0.0.0 on shared VLANs. Put TLS and an identity-aware proxy in front if you must expose it.
Verification on a clone (from advisory):
GET /setup HTTP/1.1
Host: <your-host>
Accept: application/ad+json
If the response shows invite/write: true and target at the Drive root with usagesLeft: 1, assume CVE-2026-69157 until upgraded.
Compromise indicators: unexpected Agents in write, Commits you did not author, new permission edges on sensitive resources.
What this means if you run AI coding agents
Agents are good at scaffolding self-hosted apps—docker run joepmeneer/atomic-server, compose files, Kubernetes manifests—but bad at default-deny networking:
- Review any agent commit that maps 9883/tcp publicly.
- Pin image digests after you validate a fork or patched commit (Microsoft PyPI typosquat lesson applies to image names too).
- Separate demo instances from production credentials—a “quick Atomic Data POC” on your laptop VLAN is still Drive-wide write if neighbors can reach it.
- Use Beam MCP/skill scanning before running unfamiliar setup scripts that curl
/setupfor convenience.
Beam can flag risky shell and scan MCP configs locally; it does not authenticate HTTP to your atomic-server. Network policy and human review of compose remain primary.
Atomic Data Agents and ed25519 commits
atomic-server’s security model ties writes to ed25519-signed Commits. CVE-2026-69157 abuses the invite plugin to register a new signing identity—not to break crypto primitives. That distinction matters for incident response: rotate compromised Agent keys, audit Commits for unexpected subjects, and do not assume revoking a password is enough when public-key Agents are the root of trust.
Developers experimenting with Agents as first-class actors in knowledge graphs should read this alongside token inheritance patterns for other agent platforms.
Secondary flaws named in the same advisory
The GitHub review text does not stop at invite redemption. Operators planning a fork or private patch should track:
- TOCTOU on
usagesLeft— a read-check-decrement pattern allows concurrent redemption of a single-use invite if attackers parallelize requests—worse than “one free admin” on busy LANs during hackathons. - Inverted expiry check — reporter states
if expires > now { reject }is backwards, so expired invites stay valid while valid invites fail, breaking both security and operator expectations during key rotation. - Downstream root impersonation — with Drive write, attackers may reach server root private keys (separate report in the advisory narrative), enabling Commits that appear to come from the legitimate root Agent.
None of those are separate CVE IDs in the published GHSA text Beam reviewed, but they explain why “block /setup on the firewall” is a bridge, not an architecture.
Docker and agent-generated compose patterns
A typical agent-suggested stack might map 9883:9883 and publish README instructions to “open /setup once.” That workflow is exactly what the PoC abuses:
| Pattern | Risk |
|---|---|
ports: ["9883:9883"] on shared Wi-Fi | Neighbor redeems invite before you finish README |
| No TLS on bootstrap | Passive observers capture commit traffic later |
| Latest image tag in compose | Unpinned digest hides version drift across laptops |
Hardening checklist for agent-maintained manifests:
- Bind to
127.0.0.1unless a reverse proxy handles auth. - Pin image digest after you validate a non-vulnerable build.
- Add a human gate on first commit that touches
atomic-serverservice definitions in CI. - Store ed25519 admin keys offline; compare Drive
writearrays after every deploy.
Comparison to other bootstrap URL mistakes
CVE-2026-69157 rhymes with default admin routes on self-hosted AI stacks—unauthenticated setup wizards, first-run flags, and MCP inspector proxies left open on 0.0.0.0. The atomic-server variant is sharper because cryptographic Agents are the authorization primitive: compromising /setup is equivalent to minting a root CA for the dataset, not merely reading public resources.
For coding agents, the fix is procedural: treat /setup like kubectl create clusterrolebinding—a command class that should never run from an LLM without a second human on video if the endpoint is network reachable.
Incident response if you suspect takeover
- Snapshot the Drive JSON and Commit log before taking the instance offline.
- Enumerate
writearrays on the Drive root; remove unknown Agent URLs. - Rotate every legitimate Agent key that shared the instance.
- Rebuild from a known-good image on an isolated network; do not copy
sledstores forward without forensic review. - Document which compose commit exposed the port—feeds back into CI CVE gates.
Atomic Data authorization model (why invites matter)
In Atomic Data, resources are URLs; rights are explicit arrays on parents; Agents authenticate via ed25519 signatures on Commits. Invites are the bridge that lets a stranger become an Agent without an existing trust chain—exactly like OAuth device codes or Kubernetes bootstrap tokens, except here the invite is world-readable at a predictable path.
When coding agents generate JSON-LD fixtures or seed scripts, they may recreate write=true invites on public collections for “developer convenience.” That pattern propagates into MCP filesystem servers that expose the same HTTP API to models. A model cannot exploit CVE-2026-69157 by itself, but an agent that curls /setup during a tutorial run can accidentally complete attacker redemption if a malicious neighbor is on the network.
Monitoring and detection hooks
Short-term detection ideas until a patched release exists:
- Reverse proxy access logs alerting on
GET /setupfrom non-admin CIDRs. - Periodic Drive root fetch in CI comparing
writeAgent subjects to an allowlist file committed in git. - Commit rate anomalies on the Drive root resource—mass deletes show up as signed Commits even when HTTP logs look normal.
Beam’s on-device collector can record that an agent ran docker compose up or curl against :9883; it does not parse Atomic Data semantics. Pair Beam with proxy logs when atomic-server sits behind nginx or Caddy agents configure.
Supply chain: Rust crate versus container image
Projects may depend on atomic-server as a crate in a monorepo or pull joepmeneer/atomic-server images. Version checks differ:
cargo tree -i atomic-server
versus inspecting Docker image labels. Agents updating Cargo.toml without reading advisories may bump patch versions that remain inside the 0.29.0–0.40.1 vulnerable band. Pin advisories in Dependabot/Renovate comments linking GHSA-pfwx-xjp4-qxcj so automated PRs do not merge blindly.
When to choose a different backend
If your agent knowledge base does not need Atomic Data semantics, consider whether plain git, SQLite, or document stores with mature auth reduce attack surface. CVE-2026-69157 is not an argument against linked data—it is an argument against exposed bootstrap endpoints on protocols where public-key Agents are root-of-trust.
Track ontola/atomic-server releases and rebuild images the day a fixed tag appears; until then, assume any tutorial that tells models to “visit /setup” is unsafe on untrusted networks. If you maintain internal docs agents read via RAG, delete or rewrite those setup steps so automated onboarding does not preserve a known-critical URL pattern.
Related reading
- Langflow MCP RCE and honeypot claims
- MCP security practical guide
- Check dependencies against CVEs in CI
- FastAPI code generator OpenAPI RCE
- November 2026 attacks timeline
- What is AI agent monitoring?
Frequently asked questions
What is atomic server?
atomic-server is the Rust reference server for Atomic Data—an open knowledge-graph stack. It exposes HTTP APIs for resources, Commits, and Agents identified by ed25519 keys. Many hobby and research deployments run it in Docker beside coding-agent workflows.
What does CVE-2026-69157 let an attacker do?
They can redeem /setup without logging in, create an attacker-controlled Agent, and gain write access to the entire Drive—enough to alter or delete all resources and permission edges on an unclaimed instance.
Which versions are affected?
The GitHub advisory lists 0.29.0 through 0.40.1. Check your container tag or Cargo.toml dependency before assuming a latest pull is safe.
Is there an official fix yet?
Patched versions are none in the advisory as of publication. Track ontola/atomic-server releases and rebuild images when a fixed version ships.
How do I check exposure?
From a safe network position, inspect /setup JSON for root-targeted write invites and test port bindings with Docker or Kubernetes introspection—agents should not be the only audit trail.
Why should coding-agent users care?
Agents generate compose and README quick starts that expose services by default. This CVE turns a bootstrap convenience URL into anonymous admin on fresh installs.
Sources
Advisory fields and version ranges reflect the GitHub-reviewed publication as of November 18, 2026 UTC.

