
Please share this around to prevent users from falling victim to this exploit, full findings & how it worked will be published once @telegram has patched it. We have also reported this via their official channels with all details!

Abliterated and uncensored model APIs. GLM, DeepSeek, Kimi and more. OpenAI-compatible access.

Please share this around to prevent users from falling victim to this exploit, full findings & how it worked will be published once @telegram has patched it. We have also reported this via their official channels with all details!

This Vulnerability is currently live, and we have only disclosed enough so @telegram can patch it immediately. - fix the scope-teardown lifetime bug in your QuickJS-ng fork and bump it - move widget execution into its own process, so engine bugs aren't client bugs - give that process a low-integrity or AppContainer token

It re-fires on every render, so cleaning doesn't stick — scroll back to the message and it reinstalls. Seen on Windows 11, Telegram Desktop 7.4.2 x64. The process tree, verbatim: powershell.exe spawned by Telegram.exe, with curl.exe and the Startup executable below it. No browser anywhere in the chain, which is what makes it easy to miss.

From there the script gets read/write inside the client process, finds a Windows API the executable already imports, and calls it with a PowerShell command line. Nothing has to be downloaded first and nothing needs elevation — it runs as the user. The payload in the wild is a dropper: a DLL side-loaded from a game folder so it runs inside a signed process, plus an executable in the Startup folder for persistence. Defender only flagged both after they had already run.

Root cause is a use-after-free in the engine's scope teardown. The syntax that trips it is ordinary JavaScript, no special APIs. A scope is freed while a closure created inside it is still alive, and writing through that closure afterwards lands in freed memory.

No interaction is needed. Rendering is enough: open the chat, or scroll back to the message, and the script executes. /start only matters because it makes a bot reply show up in the first place.

The vulnerability is in the new widget feature (Bot API 10.4). A bot can attach HTML and JavaScript to a normal message, and the client runs that script when the message renders inside QuickJS-ng, on a thread in the Telegram process.

Today one of our users reported getting hacked by a single @Telegram message. No link. No tap. No mini app. They pressed /start on a bot. We ran the message through our abliterated models. It's zero click code execution in Telegram Desktop. 🧵

The @Kimi_Moonshot team's decision to reduce cyber weights makes for interesting results, DeepSeek abliterated variants and GLM abliterated variants seem to lead on specific evals. More soon.

We are currently running eval benchmarks and testing abliterated weights for kimi-k3.1-pro-abliterated against other abliterated models. Full results and release details will be posted soon, Updates/Docs: huggingface.co/routerhub/kim…

Share your Routerhub link with others and get rewarded for spreading the word: routerhub.si/dashboard/refer…

Referral System is now live on routerhub.si 🫂 Earn 15% of any referred user's lifetime deposits paid out via crypto or added to your Routerhub balance instantly.

Deepseek v4.1 flash 1017 Abliterated models are now on Routerhub!
New model(s) added: [+] deepseek-v4.1-flash-1017-abliterated | $0.30 per M input [+] deepseek-v4.1-flash-abliterated | $0.30 per M input [+] deepseek-v4.1-flash-uncensored | $0.15 per M input [+] deepseek-v4-pro-abliterated | $0.40 per M input [+] deepseek-v4-pro-uncensored | $0.20 per M input routerhub.si/models

New model(s) added: [+] deepseek-v4.1-flash-1017-abliterated | $0.30 per M input [+] deepseek-v4.1-flash-abliterated | $0.30 per M input [+] deepseek-v4.1-flash-uncensored | $0.15 per M input [+] deepseek-v4-pro-abliterated | $0.40 per M input [+] deepseek-v4-pro-uncensored | $0.20 per M input routerhub.si/models

New model(s) added: [+] kimi-k3.1-pro-uncensored | $0.3 per M input [+] kimi-k3.1-pro | $0.2 per M input routerhub.si/models
Releasing the model weights of Kimi K3.1 Pro: huggingface.co/moonshotai/Ki… We'll keep advancing open intelligence.
