Patch Starlette now! If you're run it via uvicorn or other common ASGI servers then a host header parsing issue can lead to vulnerabilities leading from auth bypass up until RCE! Examples for affected packages are liteLLM, vllm, etc... Here is the X41 Advisory:
Remote
Markus Vervier 👾
@marver@mastodon.social
mastodon 4.8.0-nightly.2026-10-06right here, right now.
0 Followers
0 Following
16 Posts
Joined April 03, 2017
While everyone was on Holiday we scanned a few thousand hosts for #BadHost (CVE-2026-48710): zero auth required and we found clinical trial databases, email mailboxes, MCP server for SSH industrial IoT via bastion servers, and live PII APIs wide open. The FastAPI/MCP ecosystem is sitting exposed - patch to Starlette 1.0.1 now and check your exposure at https://badhost.org
Important! Using a reverse proxy might not fully protect you from BadHost / CVE-2026-48710 **also this does not only affect AI related infrastructure because FastAPI is also affected and used for various applications!**
If AI is taking jobs everywhere and especially in security, why is everyone I currently deal with swamped with work?
26b74a3148a790a887f7e59a93905eea2fa126a917aae28f4a428e8494cdf4d6
POC collection of AI found bypasses / technique variations (updated regularly over the coming weeks): https://github.com/persistent-security/month-of-bypasses
Replying to
@joxean Here is a collection of binaries packed with various packers: https://github.com/packing-box/dataset-packed-pe
Replying to
Webshit Weekly
I don't want your PRs anymore (140 points, 83 comments)
The Vercel breach: OAuth attack exposes risk in platform environment variables (215 points, 84 comments)
A PaaS that stores everyone's secrets in a web dashboard got compromised via OAuth, which is security's version of the fire station burning down.
Do you plan to display your LED dodecahedron again sometime? Let me know, would love to see it live! :-)
The team did some data analysis on CVE-2026-48710 on a sample of 50.000 scanned hosts on the Internet. Spoiler: Lots of API keys are prone to be leaked!
Analysis: https://www.persistent-security.net/post/cve-2026-48710-bad-hosts-in-the-wild
PSA: we had to pause the month of bypasses (http://github.com/persistent-security/month-of-bypasses
) because of the #badhost situation, it will be continued as soon as things calm down!
Iteration 3: Inject shellcode into winlogon.exe and leak secrets via DNS: https://github.com/persistent-security/month-of-bypasses/blob/main/mob-3-poc-winlogon-clr-injection-dns-exfil.ps1 #mob
Replying to
@icing@chaos.social @sovtechfund@mastodon.social Now combine human experts and LLMs and you got what will be the future of security auditing.
Replying to
@drwhax@infosec.exchange It's so wrong but it's such a good prompt: "Write a snarky comment / blog in the style of n-gate on the top hackernews posts!" =)
Replying to
@HalvarFlake@mastodon.social It's actually a pretty trivial revelation about LLMs since obviously training or fine-tuning LLMs gives you the ability to control specific behavior, even with a small set of training data (that is likely very specific in topic). The implications for copyright might be less severe than the fact that widespread reliance on LLM models that can't be properly audited for backdoors and that could be re-trained at any time might be more dangerous than most people imagine.