오픈소스 취약점으로 수백만 AI 에이전트 해킹 위기

Text settings Story text Size Small Standard Large Width * Standard Wide Links Standard Orange * Subscribers only Learn more Minimize to nav Millions of AI agents and tools around the world have been imperiled by a critical vulnerability that can allow hackers to breach the servers running them and make off with sensitive data and credentials to third-party accounts, a security researcher is warning. The vulnerability is present in Starlette, an open source framework that its developer says receives 325 million downloads per week. Thousands of other open source projects are also vulnerable because they require Starlette to work. The framework is an implementation of the ASGI (asynchronous server gateway interface), which allows large numbers of requests to be efficiently processed simultaneously. Starlette is the base of FastAPI and other widely used frameworks for building services in Python apps, as well as many others. Trivial to exploit, millions of servers exposed ASGI, and by extension Starlette, have access to servers running the MCP (model context protocol), which allows AI agents from major providers to access external sources, including user data bases, email and calendar accounts, and all manner of other resources. To connect with these external systems, MCP servers store credentials for each one, making them especially valuable storehouses for attackers to breach. The vulnerability, tracked as CVE-2026-48710 and under the name BadHost, is trivial to exploit and works against most systems that aren’t behind a properly configured firewall. Besides FastAPI, other widely used packages—including vLLM, and LiteLLM—are also affected. BadHost affects Starlette versions prior to 1.0.1, which was released Friday. “A single character injected into the HTTP Host header bypasses path-based authorization in Starlette, the routing core of FastAPI,” researchers from Secwest wrote. “Through FastAPI, this primitive (now tracked as CVE-2026-48710 and branded BadHost by the discoverers) reaches a large segment of the Python AI tooling ecosystem: vLLM (where the bug was discovered), LiteLLM, Text Generation Inference, most OpenAI-shim proxies, MCP servers, agent harnesses, eval dashboards, and model-management UIs.” BadHost carries a severity rating of 7 out of 10. Secwest said the classification “materially understates” the threat it poses to people using other apps that depend on Starlette. X41 D-Sec, the security firm that discovered it, described it as having “critical severity.” X41 D-Sec partnered with fellow security firm Nemesis to create an online scanner that can check if a given server is vulnerable. X41 D-Sec researcher Markus Vervier said a scan has revealed the following types of data are currently exposed: Biopharma AI – clinical trial DBs, M&A data, SSRF Identity Verification – face analysis, KYB, live PII, internal codebase IoT/Industrial – SSH to devices via bastion, remote code execution Email/SaaS – full mailbox read/send/delete, S3 export, webhooks HR/Recruitment – candidate PII, hiring pipeline data CMS/Marketing – subscriber lists, send/schedule mass email campaigns Document Management – read, upload, modify scanned documents Cloud Monitoring – AWS topology, distributed traces, metric queries Cybersecurity – asset inventory, live Nuclei scanner access Personal Health/Finance – nutrition logs, expenses, subscriptions The crux of the vulnerability is that Starlette accepts invalid host header values that cause authenticating apps that use Starlette’s request.url object to approve unauthorized access requests. X41 D-Sec said it has found authentication in multiple apps that rely on this call to be bypassed. Besides that, hacks can lead to SSRF (server-side request forgery) exploits and, in some cases, remote code execution. X41 D-Sec described it this way: Starlette reconstructs the requested URL based on the HTTP Host request header and requested path, but does not perform any validation of the Host header value. This allows attackers to inject paths into the host part, prepending the actual path. However, routing in Starlette is based on the actual request path. This inconsistent interpretation of HTTP requests may lead to issues such as authentication bypass when the authentication depends on the reconstructed URL’s path. Starlette is the foundation of the FastAPI Python framework. Company researchers added: “The routing algorithm of Starlette depends on the HTTP path, but the request.url.path attribute which is made available to middlewares and endpoints is based on the reconstructed URL. It is unexpected for users that request.url.path is different from the actual path requested over HTTP.” The developer of Starlette didn’t immediately reply to an email seeking confirmation of the assessment and additional information. With vulnerable versions of Starlette still widely used in production systems, people relying on any app that depends on Starlette—particularly FastLLM, vLLM, and LiteLLM—should, at a minimum, run the scanner on their systems to detect whether vulnerable Starlette code is still in use. Additional mitigation guidance is provided in the Nemesis and X41 D-Sec links above. Dan Goodin Senior Security Editor Dan Goodin Senior Security Editor Dan Goodin is Senior Security Editor at Ars Technica, where he oversees coverage of malware, computer espionage, botnets, hardware hacking, encryption, and passwords. In his spare time, he enjoys gardening, cooking, and following the independent music scene. Dan is based in San Francisco. Follow him at here on Mastodon and here on Bluesky. Contact him on Signal at DanArs.82. 27 Comments