CRITICAL
GHSASUPPLY-CHAIN
LangBot: Authenticated RCE Via MCP Configuration
Authenticated users can run arbitrary commands on the server and fully compromise the host.
2 CRIT · 6 HIGH · 1 MED · 3 INFO · THREAT RED · 12 items · Generated in 243s
Authenticated users can run arbitrary commands on the server and fully compromise the host.
Unauthenticated callers can use the operator's Meta token to read and modify Meta Ads data through exposed MCP tools.
Unauthenticated SSRF can make the MCP server reach internal hosts and cloud metadata endpoints, exposing internal services or secrets.
Attackers can read sensitive local files the MCP server can access, exposing secrets, credentials, or other users' session data.
Audio users hear can differ from model input, letting hidden or altered content bypass moderation and safety checks.
A crafted regex can hang inference workers, causing denial of service in vLLM structured output handling.
Remote clients can crash shared vLLM workers and keep inference unavailable until the service is restarted.
An oversized audio upload can force memory allocation before validation, enabling denial of service on vLLM speech endpoints.
Brief mention of AI agents exploited via WhatsApp, but no technical detail to assess impact or required defenses.
Describes a vendor defense strategy, not a specific flaw or exploit, but it may affect how AI threats are detected and remediated.
Untrusted page or thread content can steer an agent into unsafe clicks or command execution during normal work.
Highlights that AI agents change security operations, but provides no specific exploit, weakness, or affected product.