Malicious Code Injected into LiteLLM Open Source AI Package via Supply Chain Attack
The open-source software LiteLLM, designed for generative AI applications, was compromised when malicious code was injected into its official distribution packages. The attack allowed threat actors to steal authentication credentials and execute arbitrary code on affected systems.
Developers disclosed on October 21, 2026, that version LiteLLM 1.107.0 had been altered during the release process. This constituted a software supply chain attack in which attackers gained access to the build pipeline and distributed the tainted software through trusted repositories.
According to researchers at JFrog, seven malicious wheel files were published on the PyPI Python package index. Official container images hosted on Docker Hub and GitHub Container Registry were similarly affected.
The implanted malware implements credential-harvesting routines that target cloud service keys, SSH private keys, browser-stored credentials, and development environment tokens. It can also extract API keys for generative AI services directly from the LiteLLM database.
In addition, the code contains functionality to fetch and execute payloads from remote servers. It leverages Python's automatic execution hooks that run at interpreter startup, allowing the malicious logic to activate even if the LiteLLM application itself is never explicitly imported.
This is the second supply-chain compromise involving LiteLLM in 2026; developers stated that the March incident was unrelated to the current breach.
Related articles
Anthropic Funds New Security Developer Role at Python Software Foundation Following Sharp Rise in Vulnerabilities
Anthropic has financed a new security position within the Python Software Foundation in response to a marked increase in vulnerabilities and malicious packages affecting the Python ecosystem. Data published by Seth Larson showed a significant spike in reported issues during 2026. Brett Smith has been appointed as Security Developer and will collaborate with Larson, who serves as Security Developer-in-Residence, and Mike Fiedler, the PyPI security engineer. The expanded team will focus on scaling automated quarantine mechanisms in PyPI and streamlining the workflows of the Python Security Response Team. At PyCon US 2026, the team reported handling three to four times more vulnerability disclosures for Python and pip compared with the previous year, alongside a 60 percent increase in malware detections and removals from the repository.
Malicious ZIP Archive Delivers Git-Based Remote Code Execution via Hooks and Filters
A developer received a fake job offer from Cryptic Heaven through LinkedIn, leading to a ZIP file containing a rigged Git repository. The archive included a pre-configured .git directory with altered hooksPath pointing to .git/objects/interrupt and a custom clean filter named ltr. Upon git status or branch checkout, the setup triggered a shell script that downloaded and executed remote code from a server hosted on vercelscript.vercel.app. The attack leveraged post-checkout hooks and Git filters to run silently without requiring install scripts or explicit execution. Multiple trigger paths existed, including automatic detection in Visual Studio Code, Git-aware terminal prompts, and following README instructions to switch branches. The technique worked on both Linux/macOS and Windows through Git for Windows shell support. The incident highlights risks of opening untrusted Git repositories delivered outside standard clone workflows.
OIDC-Based Push Deployment on VPS Removes SSH Keys from CI While Enforcing Strict Isolation
A new self-hosted deployment system called cd enables secure continuous delivery to single Linux VPS hosts running Docker without storing any long-lived secrets in CI repositories. The architecture relies on short-lived OIDC JWT tokens issued by GitHub Actions or GitLab CI, verified on the target server through JWKS public keys. The server component cdd performs token validation, archive checks, resource-limited builds inside BuildKit, health checks, and automatic rollback on failure. All containers are launched with read-only root filesystems, dropped capabilities, no-new-privileges, and network binding restricted to 127.0.0.1 behind Caddy or Nginx. Administrative control is provided by the cdctl CLI that registers projects, sets RAM/CPU limits, and manages DNS records via Porkbun integration. The entire solution avoids external databases, web panels, and persistent secrets while enforcing one-repository-to-one-application boundaries.
Stolen Credentials Poison cPGuard Update Channel, Turning Security Software into Data Theft Backdoor
Attackers used compromised OPSSHIELD credentials to hijack the cPGuard update distribution network on November 2 at 02:06 UTC, injecting malicious scripts into versions 5.88.12 through 5.88.19. The poisoned updates targeted Linux servers running the security tool, which holds root privileges on hosting provider and cloud infrastructure machines. The script harvested /etc/shadow, /etc/passwd, MariaDB/MySQL root credentials from my.cnf, root shell history containing API tokens, and SSH authorized_keys files before exfiltrating them via the vendor's own infrastructure. A persistent backdoor was also dropped at /var/tmp/.cpg91511_sh. OPSSHIELD detected the intrusion within 12 hours and halted distribution, but confirmed that 32 customer servers had their data accessed. The incident highlights risks of relying on static credentials for security product update pipelines without code signing or behavioral restrictions.