Your favorite AI apps probably run on a piece of software called LiteLLM. It gets downloaded 97 million times a month. Yesterday, someone poisoned it.
Here's what happened. LiteLLM is like a universal adapter that lets companies plug into AI models like ChatGPT and Claude. If you've used any AI tool in the last year, there's a decent chance LiteLLM was running somewhere behind it. A hacking group called TeamPCP uploaded a fake update to the online store where developers download it. If anyone installed that update, the malware quietly scooped up every password, secret key, and login credential on their computer, packaged it, and sent it to the hackers. You didn't even need to download LiteLLM yourself. If you downloaded any app that uses LiteLLM in the background, the same thing. You got hit without knowing LiteLLM existed.
The only reason anyone caught it is that the malware had a bug. One developer's computer ran out of memory and crashed. That crash is the only reason the attack got noticed. Without that bug, this could've run for weeks.
But the scariest part is how the hackers got in. Five days before, TeamPCP broke into a security tool called Trivy. Trivy's job is literally to scan code and find weaknesses. They turned the guard dog into the burglar. Then they used stolen passwords from that hack to break into another security company. Then they used passwords from that hack to steal the login that lets someone publish new versions of LiteLLM, because LiteLLM used Trivy in its own security checks. The lock on the front door was the thing that was compromised.
When developers tried to report the break-in on GitHub (where the code lives), the hackers used the stolen account to shut down the report. Then they flooded it with 88 fake comments from 73 hijacked accounts in 102 seconds to drown out anyone trying to warn people.
In five days, TeamPCP broke into five different software platforms. Each break-in gave them the password to the next one. A domino chain. The poisoned update was live for about 3 hours before it got pulled. LiteLLM gets 3.4 million downloads per day. And TeamPCP posted on Telegram, saying more attacks on open-source tools are coming in the months ahead.
The entry point was the security scanner that was supposed to keep LiteLLM safe. That's the part I keep coming back to.

X (formerly Twitter)
Andrej Karpathy (@karpathy) on X
Software horror: litellm PyPI supply chain attack.
Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds,...