← All tools
IDEs & agents · LibreChat (open source)

LibreChat

Open-source, self-hosted ChatGPT-style chat app that connects to many AI providers and supports agents, actions, MCP servers and multi-user sign-in.

3Advisories
0critical
2high
1medium
2026-11-20Latest
Timeline · advisories per quarter
32026 Q4
Named in sources

Reported for LibreChat

01. highSecret exposureNew this weekCVE-2026-54044: In LibreChat, a collaborator with EDIT access to a shared agent could repoint an action to their own server and receive the owner's stored API key (fixed 0.8.7)LibreChat agents can carry Actions: OpenAPI-described tools whose API key (or OAuth client secret) the agent owner stores encrypted and LibreChat sends as a Bearer token whenever the model calls the tool. Before 0.8.7, POST /api/agents/actions/:agent_id merged an editor's new metadata.domain and metadata.raw_spec into the existing action but kept the owner's api_key and OAuth client secrets. A collaborator the owner had shared the agent with at EDIT permission could swap the spec and domain to a host they control, and the next time the agent ran the tool LibreChat decrypted the owner's credential and sent it there. Fixed in 0.8.7. LibreChat OpenAI 2026-11-20 02. highBroken access controlNew this weekCVE-2026-54039: LibreChat let an unauthenticated attacker pre-register a victim's email with a forged social provider and share the account once the victim signed in with Google or another provider (fixed 0.8.6)From 0.7.0 until 0.8.6, when local registration is enabled and at least one social login provider is configured, LibreChat's POST /api/auth/register wrote the provider field from the request body onto the new user record. The social-login email fallback later accepted that record when the victim signed in, because it compared only the stored provider name and not the provider's subject id. An unauthenticated attacker who knows a victim's email could pre-register it with provider set to, for example, google and a password of their choice; after the victim's first OAuth sign-in both sessions are bound to the same user, giving the attacker lasting read and write access to the victim's conversations and data. Fixed in 0.8.6. LibreChat 2026-11-20 03. mediumAgent hijackingNew this weekLibreChat: a malicious remote MCP server could run script via its description, MCP OAuth followed redirects to internal URLs, and another user's file could be pulled into the model context, plus 7 more CVEs (CVE-2026-54023 to 54046; fixed 0.8.6–0.8.9)Alongside the two LibreChat CVEs covered earlier, NVD published ten more on 20 November 2026 (GitHub advisories from 17 November). The AI-specific ones: MCPConfigDialog rendered a remote MCP server's description through dangerouslySetInnerHTML, so a malicious MCP server could run script in the LibreChat origin when a user opened its config dialog (CVE-2026-54032, fixed 0.8.6); MCP OAuth discovery, registration, token, refresh and revocation requests validated only the first URL and then followed redirects to internal addresses (CVE-2026-54031, fixed 0.8.8); and chat attachments accepted another user's file_id, putting that user's extracted document text into the attacker's model context (CVE-2026-54043, fixed 0.8.7). 0.8.5 also sent federated OAuth access, ID and refresh tokens to the browser via GET /api/user (CVE-2026-54023). Upgrade to 0.8.9 to get all fixes. LibreChat MCP SDKs & servers OpenAI 2026-11-20
The weekly email

One email a week. Only what changed.

Every new security hole in AI coding tools, MCP servers, app builders and the stacks they generate (Next.js, Supabase, Node, PHP), each with how to check your app and the fix. No roundups of things you already read.

Free, one email a week. Read the latest issue first if you like. One-click unsubscribe in every email. We never sell or share addresses.