The security incident, which took place on October 21, specifically impacted version 1.107.0 of the widely used software. The rogue packages remained live and accessible to the public for roughly three hours before being detected and neutralized. According to data analysis and download tracking estimates, the poisoned packages were downloaded over 49,000 times during this narrow window of exposure.
In response to the compromise, maintainers and security teams swiftly removed the malicious releases from PyPI. Simultaneously, Docker Hub tags and image variants associated with the affected version were either completely removed or promptly repointed to clean, verified artifacts to halt any further downstream distribution.
According to a detailed post-incident investigation published by LiteLLM, the attackers managed to breach the infrastructure by obtaining GitHub CLI OAuth tokens. These sensitive credentials belonged to an unexpected, innocuously named user account operating within an internal LiteLLM deployment that had been inadvertently exposed to the public internet.
Investigators revealed that these tokens were likely leaked as a byproduct of using Claude Code. Under specific operational circumstances, the AI coding assistant would embed such authentication tokens directly into a response body, which was subsequently captured and recorded in LiteLLM internal system logs. The attackers appear to have successfully located and extracted the tokens from these exposed log files.
Despite uncovering how the tokens were leaked, LiteLLM noted that it remains entirely unclear how the external attackers initially gained unauthorized access to the internal deployment or managed to move laterally within the network. At the time of the breach, the threat actors gained access to three separate sets of developer GitHub CLI credentials.
Armed with these compromised credentials, the attackers moved quickly to subvert the software release pipeline. They bypassed standard security controls by disabling branch protections within the repository settings. Following this, they merged a malicious change that redirected the software releaser to an entirely different source repository. Finally, they executed a compromised GitHub workflow pointing to a lookalike repository, allowing them to digitally sign and publish the poisoned software artifacts to the public package repositories.
Users who inadvertently downloaded and installed the compromised LiteLLM version 1.107.0 packages during the brief compromise window found their systems infected with a multifaceted payload. This malware functioned simultaneously as an advanced information stealer and a remote access trojan (RAT). LiteLLM confirmed in its disclosures that enterprise customers who pull updates exclusively from dedicated private repositories remained unaffected by the attack, limiting the blast radius primarily to public open-source users.
The malicious package appeared to have been constructed with the assistance of artificial intelligence tools and was served from a lookalike repository named Berri-AI/litellm. This malicious namespace was carefully crafted to impersonate the legitimate Berri-AI/litellm repository managed by the project. Upon execution, the payload automatically extracted three distinct files directly into the root of the Python site-packages directory. This clever persistence mechanism ensured that the malicious code would be loaded and executed automatically every time a local Python interpreter was initiated on the host system.
Once activated, the interpreter first loads the information stealer component of the malware. This module is designed to comprehensively comb the host system for valuable data, harvesting broad system information, stored browser credentials, session cookies, developer authentication tokens, and local environment variables.
Furthermore, the stealer targets high-value enterprise and cloud assets, extracting sensitive cloud secrets associated with major infrastructure providers including Amazon Web Services, Google Cloud Platform, and Microsoft Azure. It also searches for and exfiltrates proprietary AI tool API keys belonging to platforms such as Anthropic, GitHub Copilot, and OpenAI, alongside local SSH private keys, shell history files, and GPG encryption keys. Once all target data is gathered, the malware encrypts the payload and exfiltrates it to an external domain controlled by the attackers, identified as litellm[.]dev.
Cybersecurity firm Wiz reported that the attackers wasted no time leveraging their spoils, beginning to abuse the freshly stolen credentials roughly four and a half hours after the initial data exfiltration occurred. Wiz researchers observed the threat actors actively querying Amazon Bedrock models utilizing the newly acquired AWS credentials, indicating a classic case of LLMjacking, where compromised cloud environments are hijacked to run unauthorized, resource-intensive artificial intelligence workloads at the expense of the victim.
In addition to the credentials harvesting and cloud abuse, security firm StepSecurity analyzed the remote access trojan component of the malware. Researchers observed the RAT maintaining regular contact with the attackers’ command-and-control infrastructure. According to StepSecurity’s technical analysis, the malware also checks a GitHub dead drop on an hourly basis, downloading and executing arbitrary Python code provided directly by the attacker in real time.
Security experts and maintainers strongly advise any users who installed LiteLLM version 1.107.0 to take immediate remedial action. Affected individuals and organizations must remove the compromised version immediately, rotate all exposed developer, cloud, and API credentials, and perform thorough forensic sweeps of their systems to check for persistent malware artifacts. The project has confirmed that the most recent clean and secure version available for deployment is LiteLLM 1.106.2.

