Two LiteLLM Attacks and the Checkpoint Your Pipelines Need LiteLLM suffered two separate supply-chain compromises in 2026, the first in March when attackers bypassed official CI/CD workflows to publish malicious PyPI versions carrying a credential stealer that harvested environment variables, SSH keys, cloud credentials, Kubernetes tokens and database passwords, and the second in October when an administrative key committed to a public repository in May and never revoked let an attacker extract a GitHub organization owner's token and two maintainer tokens. CloudSEK reported in August that the March attack had exposed more than 2,500 organizations and about 434,000 CI/CD pipelines, after which LiteLLM's daily PyPI downloads dropped sharply and stayed at that lower level. Anaconda is pitching its curated AI Artifacts channels, which apply a build delay between upstream source availability and build initiation, as a checkpoint between upstream releases and enterprise pipelines. In the age of agentic development, every day it seems a major cybersecurity incident takes place. One well-known incident was the LiteLLM supply chain attack in March 2026. According to LiteLLM, initial evidence suggests the attacker got malicious versions of LiteLLM onto PyPI by “ bypass\ ing\ official CI/CD workflows https://docs.litellm.ai/blog/security-update-march-2026 what-happened ” within LiteLLM’s normal release flow. LiteLLM said the malicious versions included a credential stealer designed to harvest all sorts of secrets and credentials across the enterprise landscape, such as environment variables, SSH keys, cloud credentials, Kubernetes tokens, and database passwords https://docs.litellm.ai/blog/security-update-march-2026 what-happened . If this were with a small Python library, this wouldn’t be as big of a deal, but LiteLLM has millions of monthly downloads from PyPI and billions of total downloads. After this incident, LiteLLM revamped its CI/CD security with additional security scans, new Docker image protocols, and other adds https://docs.litellm.ai/blog/ci-cd-v2-improvements . LiteLLM downloads stayed high through mid-August 2026. In August, CloudSEK reported that the March attack had exposed more than 2,500 organizations and about 434,000 CI/CD pipelines https://www.cloudsek.com/blog/ai-supply-chain-breach-2500-companies-434000-cicd-pipelines . Shortly after this news came out, daily downloads from PyPI dropped precipitously, and have stayed at that lower level ever since https://pypistats.org/packages/litellm . One possibility is that many security teams stopped letting pipelines pull new releases straight from PyPI. Those security teams had reason to be cautious. In October, LiteLLM was compromised a second time, through a different path. An administrative key committed to a public repository in May was never revoked. In October, the attacker used it to extract a GitHub organization owner’s token, then pulled two more maintainer tokens out of CI secrets. LiteLLM’s branch protection rejected the attacker’s change to the release workflow, but the owner’s token could override it. That same week, LiteLLM had relaxed its release approval, normally tied to one device in its office, so any two engineers could trigger and approve a release. The controls existed. One over-privileged credential that was never rotated was enough to get around them https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain . Put a Checkpoint Between Upstream Releases and Your Pipelines If even well-defended projects can be compromised, the safeguard has to sit between upstream releases and your pipelines. In between these attacks enterprises either removed LiteLLM completely from their repos, used older versions released before the compromise, or they switched their open source LiteLLM processes to an enterprise-grade secure version of the framework, such as Anaconda’s curated package channels, part of Anaconda’s AI Artifacts https://www.anaconda.com/platform/ai-artifacts . Pinning versions and mirroring known-good releases both work, but only for the packages you’ve already identified, and they have to be maintained by hand for each one. A typical enterprise environment depends on hundreds of packages, any of which could be next. What enterprises might not know is how Anaconda keeps compromised releases from reaching its enterprise-grade 20,000+ Python package repository in the first place. What keeps a compromised release out of your environment? The answer comes from the list of secure build practices Anaconda uses for its repository. In both attacks, the malicious versions were live on PyPI for only hours before they were caught. The first defense is not building a new version the moment it appears. When Anaconda builds a new version of a package such as LiteLLM, it uses a build delay period, or “ a time delay between source availability and build initiation, providing a window to detect upstream compromises before packages enter the build pipeline https://www.anaconda.com/docs/reference/policies-practices/security .” This build delay is designed to keep malicious releases like these out of the repository. The October release was live on PyPI for about two hours, which is plenty of time for automated pipelines to pull it. Anaconda sets the delay for each package, so your organization doesn’t have to spend cycles working out what it should be. The second defense is building from source. Anaconda builds most packages from source code in its own controlled build infrastructure rather than redistributing prebuilt packages https://www.anaconda.com/docs/reference/policies-practices/security . Prebuilt packages come out of the upstream project’s own release pipeline, which is exactly what attackers took over in October. When Anaconda builds the package itself, a compromised upstream build doesn’t become the package your pipelines install, and Anaconda can see what goes into each package https://www.anaconda.com/docs/reference/policies-practices/security . Building from source won’t catch tampering in the source code itself, which is why the build delay comes first. Anaconda’s channel went from version 1.82.4, built March 22, to 1.83.7, built April 30, and never carried the compromised versions from March 1.82.7 & 1.82.8 into the repository. The same can be said about October’s compromised LiteLLM version 1.107.0, as Anaconda’s channel went from 1.104.0 on October 6 to 1.108.0 on November 6 skipping the compromised version, which Anaconda never built https://anaconda.org/channels/anaconda/packages/litellm/files?sort=-updated at . These dates show that Anaconda never released the compromised versions of LiteLLM to the enterprise repository, giving organizations a safer, curated repository to pull updates and packages from. October also shows why building from source matters: independent analyses found the malware only in LiteLLM’s prebuilt wheels, not in its source code https://www.stepsecurity.io/blog/litellm-1-107-0-compromised-pypi-backdoored-wheels-credential-stealer , which is what Anaconda builds from. For your team, that means a malicious upstream release that’s caught and pulled within hours does not reach the packages your pipelines install. How Anaconda Locks Down Its Build and Release Systems Anaconda’s build and release systems are locked down by design. Only authorized staff can reach Anaconda’s build and release systems, and access requires multi-factor authentication and is limited to what each role needs. Only the authorized specialists can change a package recipe, and every change needs a teammate’s review before it merges https://www.anaconda.com/docs/reference/policies-practices/security . At Anaconda, access to build infrastructure doesn’t grant the ability to publish. Build and release use separate credentials, and only designated members of the package building team can publish. Anaconda also verifies source code against recorded SHA-256 checksums and pins versions in its build recipes. If the source has changed since it was committed, the checksum won’t match and the build stops https://www.anaconda.com/docs/reference/policies-practices/security . In plain terms, there is no single point of failure. A stolen build credential can’t publish a package. The people who can publish can’t change what goes into a package without a teammate’s review. And if the source file changes, the build stops. The CEO of Anaconda doesn’t even have access to these build systems, demonstrating the level of access control that Anaconda has with its repository build systems. These controls target the failure behind the October attack: a single compromised credential being enough to publish. No single control catches everything. A build delay works best against compromises that are found quickly, as both LiteLLM attacks were. Malicious code that stays hidden longer needs other defenses, which is why Anaconda also scans packages for malware before distribution and can pull a package from its channels through a formal removal process if a problem is found after release. These practices aren’t specific to LiteLLM. They apply across Anaconda’s channel, so the same protections cover the hundreds of other packages your environment depends on. This is a key reason why 95% of the Fortune 500 use Anaconda as their trusted source for Python. Anaconda’s security documentation https://www.anaconda.com/docs/reference/policies-practices/security covers these build practices in more detail. Your developers can keep using LiteLLM and the rest of the Python ecosystem without every upstream release becoming an incident for your security team. To see how Anaconda’s AI Artifacts can secure your pipelines, talk to Anaconda’s sales team https://www.anaconda.com/contact/sales .