An AI Coding Agent, Wide Open on 0.0.0.0, With `bypassPermissions` Already On
CVE-2026-111322 lands another unauthenticated-RCE CVE on a self-hosted AI coding agent — this time the agent's own permission layer is turned off at the factory
On October 15, CVE-2026-111322 was published against coleam00/Archon, an open-source AI coding agent / orchestrator. CVSS 3.1 9.8 Critical, CVSS 4.0 9.3, CWE-306 (Missing Authentication for Critical Function). Affected versions: 0.2.13 through 0.11.1 — eleven releases. VulDB opened an entry on Oct 16 and the CNA record is still awaiting NVD analysis.
The bug in one paragraph
Archon exposes a /api/* control API that any client uses to send messages to the AI coding agent — the same way a developer would. Per the CNA description surfaced on Strix and OffSeq Radar, that API binds to 0.0.0.0 by default and has no authentication: no token, no basic auth, no origin check in the published description. And the agent sitting behind it runs with bypassPermissions — a shipped-default flag that tells the agent not to prompt or gate the commands it receives. Line those three up and the chain is complete: anyone on the network who can reach the port sends an agent message, the agent runs it as a system command, and the attacker inherits whatever privileges the Archon service account has on the box. One unauthenticated HTTP request, RCE as the Archon user.
Why bypassPermissions is the part that matters
The authentication gap is familiar — the digest has been logging it, with minor variations, every week since September across named MCP-server implementations (mysql_mcp_server CVE-2026-59971, IBM mcp-context-forge CVE-2026-53710, PraisonAI CVE-2026-57127, Labs64 netlicensing-mcp CVE-2026-54446, Context7 CVE-2026-75130, and more). The design pattern is: ship a control plane that assumes the operator's network is trusted, let the operator opt into auth later. The LAN assumption is wrong the moment the service runs inside Docker, on a cloud VM, on a developer's laptop at a coworking space, or anywhere a reachable network attacker exists.
What's different in CVE-2026-111322 is the second half. bypassPermissions is the vendor-shipped answer to a question that is supposed to be the last line of defense for an AI coding agent: before you run this command, do you want the user to approve it? Archon's shipped default is "no." That means the authentication gap on /api/* doesn't just leak a prompt stream to an attacker — it hands them an agent that is already configured to not ask the user before running what it is told. The permission boundary that keeps an interactive agent from being a remote shell is turned off, in the box.
That pattern — the agent's own approval boundary is nominal — is the same shape Mervin Praison's AI-agent framework shipped in September as CVE-2026-57137 (approval callback fires after execution, so the tool is already gone by the time the boundary is checked). CVE-2026-111322 is the other half: approval doesn't fire at all by default. Both tell the same story — in an AI coding agent, the "do you want me to run this?" moment is where the attack surface really lives, and vendors keep shipping that moment turned off.
This is the second named CVE on this project in five months
The first was CVE-2025-69443, published May 14, 2026 by OX Security — CVSS 6.3, an unauthenticated web-to-client CORS-bypass attack that could steal API keys and issue prompts on behalf of a victim who opened a crafted HTML page. That one was found by an outside security firm; this one has no credited discoverer in any public source we could find as of today. The project's GitHub Security Advisories page lists zero published advisories. KEV: not listed. Public PoC: none indexed per VulDB. EPSS: awaiting scoring. Fixed version: none named in the CNA record.
What to do about it if you run Archon
The two credible writeups (Strix and TheHackerWire) agree on three interim mitigations because there is no vendor patch to cite yet:
- Firewall the API port so only trusted hosts reach it. If this is on a shared VM or inside a Docker host, that means a real network policy, not just "nobody's scanning my LAN."
- Rebind the service to
127.0.0.1or a specific internal interface, instead of0.0.0.0. Loopback is the right default for a dev-time service that only its operator should talk to. - Front
/api/*with a WAF or API gateway that blocks unauthenticated requests. This is the stopgap for deployments that can't rebind to loopback — not a permanent trust boundary.
None of those closes the second half of the bug. bypassPermissions is still on the agent even after the network boundary is tight. If you run Archon in a context where you want the agent to confirm actions before executing them, treat the permission posture as its own configuration surface — not something the network sandbox handles for you.
The takeaway
The self-hosted AI-coding-agent runtime tier is shipping the same design pattern as the MCP-server tier that preceded it: a control plane bound to all interfaces, no authentication in the box, and — on this one — the agent's own approval boundary turned off at the factory. The attacker doesn't need a prompt injection, a jailbreak, or any model behavior at all. They need the port, and the agent does the rest.
Watch what the agent runs. That is the boundary that stayed open here.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.