Skip to content
Technwz transparent logo Technwz

AI, Cybersecurity, Gaming, Business and Tech News

Technwz transparent logo Technwz

AI, Cybersecurity, Gaming, Business and Tech News

  • News
  • AI
  • Cybersecurity
  • Tech
    • How-to
  • Gaming
  • Business
    • Startups
  • Marketing
    • Social Media
  • News
  • AI
  • Cybersecurity
  • Tech
    • How-to
  • Gaming
  • Business
    • Startups
  • Marketing
    • Social Media
Padlock beside a laptop showing code in a server room, illustrating the LiteLLM supply chain attack
News

LiteLLM Hacked: Malicious 1.107.0 Release Steals Cloud and AI API Keys

By Technwz Editorial Team
October 23, 2026 at 10:11 AM 6 Min Read
Updated on October 27, 2026 at 10:15 AM

Microsoft’s security team says the malware gets its command server address from a file on GitHub, so attackers can switch servers without pushing a new release. Microsoft also warns that uninstalling the package is not enough, because the malicious startup files can stay behind. Affected machines should be rebuilt.

Attackers broke into the release pipeline of LiteLLM, a popular open-source tool that lets developers call AI models from OpenAI, Anthropic, Google and others through one interface. On October 21, they used it to publish a poisoned version, 1.107.0, to PyPI, Docker Hub and GitHub’s container registry.

The malicious build carried a credential stealer and a backdoor. Anyone who installed it could have leaked cloud keys, AI API keys, browser data and other secrets. The bad release was live for only a few hours, but LiteLLM is pulled about 92 million times a month from PyPI, so even a short window matters. In under two hours on PyPI and about two and a half hours on Docker Hub online, the malicious version was pulled tens of thousands of times.

In its official incident report, LiteLLM says the bad release has been pulled and every earlier version is safe. It tells affected users to roll back to 1.106.2 and rotate every exposed secret. If 1.107.0 ever ran on a machine with internet access, LiteLLM says to assume the theft succeeded.

Which versions are affected

Only LiteLLM 1.107.0 is malicious. If you pinned an earlier version, you were not exposed through this release. All times below are October 21, 2026, UTC.

WhereAffected versions or tagsExposure window (UTC)
PyPIlitellm 1.107.0 (all seven wheels)about 11:39 to 13:23
Docker Hubv1.107.0, latest, main-latest, main-stableabout 11:46 to 14:20
GitHub Container Registryv1.107.0, latest, main-latest, main-stable (litellm, litellm-database, litellm-non_root images)until about 13:51

LiteLLM says the official v1.107.0 source tag on GitHub is clean. Security researchers note, however, that the 1.107.0 source tarball on PyPI contained a tampered build config, so avoid that version entirely. LiteLLM says customers on its private enterprise edition were not affected.

Anyone who ran an unpinned pip install litellm or pulled the latest Docker tag during those windows should assume they got the bad build.

What the malware does

The attackers did not touch LiteLLM’s actual code. Instead, they slipped three extra files into the package: cook.pth, cook.py and sitecustomize.py. Python loads these automatically every time it starts, so the malware runs even before LiteLLM is imported. Simply running python or pip on an infected machine to check things out can trigger it.

Once running, it does two jobs:

  • Steals secrets. It grabs cloud and deployment credentials, AI provider API keys, saved LiteLLM proxy credentials, SSH and other private keys, environment variables, shell history, and data from Chrome and Firefox. It packs everything into a compressed archive and sends it to litellm[.]dev, a lookalike domain made to blend in. Microsoft found that the data is encrypted before it is sent and that the malware also tries to read LiteLLM proxy database tables. The background loader gets its command server address from a file hosted on GitHub, so the attackers can change servers without publishing a new package.
  • Opens a backdoor. On Linux and macOS, a background process checks in every hour for a command server address and runs whatever Python code that server sends back. That gives the attacker ongoing remote access.

The stolen keys were put to use quickly. Wiz researchers saw the attacker testing stolen AWS credentials about four and a half hours after they were taken, then checking which Amazon Bedrock AI models those accounts could reach. That points to “LLMjacking”, where criminals run up AI bills on someone else’s cloud account.

How the attackers got in

LiteLLM first suspected the attack started with an internal deployment left reachable from the internet, but its final attack-chain report tells a different story. On May 27, a config file holding the master key for LiteLLM’s internal Agent Platform was committed to a public GitHub repository. The commit was later overwritten, but the key itself was never revoked. On October 10, the attacker used that key to call the Agent Platform API and pulled out a GitHub token belonging to a LiteLLM organization owner. That replaces LiteLLM’s earlier theory that a recently patched vulnerability exposed the deployment, according to its attack-chain report.

That same night, the attacker used the owner’s token to run throwaway CI workflows that leaked repository secrets, including GitHub tokens for two maintainers. That gave the attacker working GitHub access for three LiteLLM accounts, 11 days before the malicious release went out.

The timing made things worse. Two days earlier, LiteLLM had temporarily widened release permissions so engineers could ship daily bug fixes. With a token belonging to a GitHub organization owner, the attacker turned off branch protections and pushed malicious changes. Then they ran the official release workflow from a lookalike repository, Berri-AI/litellm instead of BerriAI/litellm, so the poisoned build came out signed and looked legitimate. A second stolen account approved the production release, and a third deleted workflow records afterward, which has made the investigation harder.

This is LiteLLM’s second supply chain incident this year, after a compromise in March. Socket found no clear link between the two, though it has not ruled out the same group using new tools.

What to do now

If you use LiteLLM anywhere, including in CI pipelines, Docker builds or Kubernetes, check today. Here is the short version of the advice from LiteLLM’s incident report and the security firms tracking the attack.

  1. Check your version without starting Python. Because the malware runs whenever Python launches, look for the files directly. Search your site-packages folders and container images for cook.pth, cook.py, an unexpected sitecustomize.py, the lock file /tmp/remote_code_loader.lock, or the marker file /tmp/.pc_once_72129118c59f5636c0bb. Microsoft says the malware creates that marker file after it sends stolen data, so finding it means your secrets were likely already taken.
  2. Isolate anything infected. Take affected machines, containers and CI runners offline. Deleting the LiteLLM folder is not enough, and Microsoft warns that the malicious startup files sit outside the package and can survive an uninstall. Remove the cook* and sitecustomize* files and their cached .pyc copies, and stop the hidden background Python process.
  3. Rebuild from a clean state. Restore from known-good backups or rebuild images from scratch, pinned to LiteLLM 1.106.2, the last clean release. Pin it by digest, not by a moving tag like latest.
  4. Rotate every secret the machine could reach. That includes cloud keys (AWS, Google Cloud, Azure), AI provider API keys, GitHub and CI tokens, Kubernetes credentials, SSH keys and database passwords. LiteLLM proxy users should also rotate the provider keys stored in the proxy database, master and virtual keys, and the DATABASE_URL password. Do it from a clean device.
  5. Block the known indicators. Block the domain litellm[.]dev and the attacker IPs 168.100.8[.]57 and 165.227.201[.]8, then review logs for traffic to them. Also look for requests to raw.githubusercontent[.]com/letters9813/pythonclient-test/main/domain.txt, a GitHub file the malware reads to find its command server.
  6. Watch your cloud bills and AI usage. In AWS CloudTrail, LiteLLM suggests looking for GetCallerIdentity calls around the time you installed 1.107.0, and for your credentials being used from outside your network. Also watch for unfamiliar Amazon Bedrock calls, new IAM users or keys, and spikes in AI spend.

What we still don’t know

LiteLLM says its investigation is ongoing. It has not said how many systems were infected, though SecurityWeek estimates the bad 1.107.0 build was downloaded more than 49,000 times from PyPI before it was pulled. LiteLLM also has not said how many systems were infected or how many credentials were stolen, and it cannot yet rule out further access to the affected GitHub accounts. LiteLLM says it is now revoking every token tied to its internal systems and vendor platforms. Its postmortem says the attacker also took tokens belonging to two maintainers and several internal systems, with a focus on upstream AI inference API keys.

Microsoft has not seen the malware download a second-stage payload, but the GitHub file it reads for instructions means the attackers could still change its behavior.

Why this matters

Tools like LiteLLM sit at the center of AI setups, which means they handle the most valuable keys a company has. This attack also shows a new risk: secrets that AI coding tools pass around can end up in logs, where attackers can find them.

It has been a rough week for developers. LiteLLM’s breach landed the same day as the emergency patches for PHP2Shell, the critical PHP flaws we covered yesterday. If you run your own servers, check both today.

We will update this story as LiteLLM shares more.

Sources: LiteLLM, BleepingComputer, The Hacker News, Microsoft Security Blog

Tags:

AI securityCybersecurity / LiteLLMNewsPythonsupply chain attack
Author

Technwz Editorial Team

The Technwz editorial team covers the tools, platforms, and decisions that matter to small business owners, developers, gamers, and digital marketers. We research hosting and cybersecurity services; break down business and marketing software; and keep tabs on the gaming industry, testing what we can, cutting through vendor marketing where we can't, and writing it all up in plain language.

Follow Me
Other Articles
PHP2Shell critical PHP vulnerability warning
Previous

PHP2Shell: Critical PHP Flaws Are Under Attack, and Most Websites Need to Update Now

A human hand reaches for a red emergency stop button while an AI agent robot works at a laptop, with screens showing tasks like sending emails and making purchases
Next

Test an AI Agent’s Stop Switch Before It Acts

Search

Follow us on

Facebook X Instagram LinkedIn

Submit Your Startup

Building something new in tech, AI or cybersecurity? Share your startup with the Technwz editorial team for a chance to be featured.

Submit Now

Technwz

Technwz is a digital publication covering technology, cybersecurity, AI, business, marketing, and gaming. We publish in-depth guides, tool reviews, and industry insights, keeping you ahead in an always-changing digital world.

Quick Links

  • About Us
  • Contact Us
  • Write For Us
  • Submit Your Startup

Popular Categories

  • Tech
  • Cybersecurity
  • Business
  • Gaming
Copyright 2026 Technwz. All rights reserved.
Privacy Policy · Editorial Policy · Corrections Policy