Join our Newsletter — 33% off our NHI Course
Home› NHI Breaches› LiteLLM Supply Chain Attack 2026: How Stolen GitHub…
Breach analysis Incident: 21 Oct 2026

LiteLLM Supply Chain Attack 2026: How Stolen GitHub OAuth Tokens Shipped a Credential Stealer to PyPI and Docker

PDF for newsletter subscribers · free to read online

← All NHI breaches
By NHI Mgmt Group, edited by Lalit Choda Updated 26 October 2026 13 min read
On this page

On 21 October 2026 a malicious LiteLLM 1.107.0 was published to PyPI and to LiteLLM's Docker images. It was the second supply chain attack on the popular open source LLM gateway in 2026, and it is a different incident from the March compromise and the September MCP authentication bypass. This time the attacker did not steal a publishing token. According to LiteLLM, they held GitHub CLI OAuth tokens belonging to three engineers, used them to override branch protection, approve a production release and delete workflow records, and so shipped signed, poisoned artifacts through LiteLLM's own pipeline. LiteLLM believes the tokens were read from logs on an internet-exposed internal deployment, though how that deployment was reached is unconfirmed. The payload stole cloud, developer and AI tool credentials, and Wiz saw stolen AWS keys tested against Amazon Bedrock within hours. The packages were pulled the same day.

Key takeaways

  • Malicious litellm==1.107.0 was live on PyPI from 11:39 to about 13:23 UTC on 21 October 2026, and Docker tags (v1.107.0, latest, main-latest, main-stable) served it until 14:20 UTC, according to LiteLLM. Wiz counted about 49,000 downloads.
  • The attacker used GitHub CLI OAuth tokens from three LiteLLM engineers: one to change the release workflow, one to pass the production approval gate and one to delete workflow records, LiteLLM says.
  • LiteLLM found that Claude Code can embed a user's GitHub CLI OAuth token in a response body, so logged response bodies put the tokens into LiteLLM logs on an internal deployment that was exposed to the internet.
  • The stealer took cloud credentials, SSH keys, AI tool auth files and LLM provider keys stored in LiteLLM Proxy, and Wiz observed stolen AWS role credentials being tested against Bedrock about four and a half hours later.
  • The identity lesson: developer OAuth tokens that can approve releases are production identities, and any log that captures them becomes a secret store that needs the same protection.

At a glance

OrganisationBerriAI (maintainer of LiteLLM); developers, CI runners, Kubernetes nodes and AI applications that installed LiteLLM 1.107.0 or pulled the affected Docker tags
WhenMalicious release published and removed on 21 October 2026; disclosed by LiteLLM the same day
AttackerUnattributed. LiteLLM, Wiz and Aikido have not named an actor
Entry pointGitHub CLI OAuth tokens for three engineers, which LiteLLM believes were read from logs on an internet-exposed internal LiteLLM deployment (access route unconfirmed)
Identities abusedDeveloper GitHub CLI OAuth tokens, including one belonging to a GitHub organisation owner; the release pipeline's signing and publishing identity; downstream cloud, AI provider and AI tool credentials
ImpactAbout 49,000 downloads of a credential stealer and remote access trojan (Wiz); stolen AWS credentials tested against Amazon Bedrock; number of exfiltrated environments not disclosed
CategoryNHI; LLM / AI platform. Incident class: confirmed NHI breach (stolen OAuth tokens used to publish a malicious release)

What happened

LiteLLM is an open source Python library and proxy that gives applications one interface to many large language model providers. Because it often holds the API keys for OpenAI, Anthropic, AWS Bedrock and others, it is an attractive target. In March 2026 its PyPI publishing token leaked through a compromised Trivy scanner in its CI pipeline, which we cover in our LiteLLM PyPI package breach page. Separately, CISA listed an exploited MCP authentication bypass in September, covered in our LiteLLM MCP auth bypass page. LiteLLM says that after March it moved releases to a private repository, adopted trusted publishing, made Docker tags immutable and signed its images. The October attack went around those controls rather than through them, because the attacker acted as trusted engineers.

LiteLLM's incident post gives a minute-by-minute account. At 10:08 UTC on 21 October, GitHub branch protection rejected an unauthorised change to the private release workflow. Five seconds later a privileged developer credential was used to override the protection and push the change, which pointed the release job at a lookalike repository, Berri-AI/litellm. Wiz says the attacker had pushed ten commits to that repository between 08:56 and 10:46 UTC, nine of them GitHub-signed. A staging run completed at 11:08, a second developer credential satisfied the production approval gate at 11:11, and seven malicious wheels reached PyPI between 11:39 and 11:40. Container jobs followed at 11:46. At 11:53 a third developer credential was used to delete the production and staging workflow records, which limited the logs available to investigators. LiteLLM also says release permissions had been widened the previous Monday so two engineers could publish bugfix backports.

How the attacker got the tokens is the identity heart of the story. LiteLLM says its team found that Claude Code can embed a user's GitHub CLI OAuth token in a response body, and that where response bodies are logged, those tokens can end up in LiteLLM's own logs. It found an unexpected, innocuously named user account on an internal LiteLLM deployment that was reachable from the internet, and believes the attacker used it to read the tokens from the logs. LiteLLM had not confirmed how the attacker got into that deployment but suspected a recently patched vulnerability. A note added on 23 October says some earlier assumptions in the post were inaccurate or incomplete and points to a separate record, so details may still change.

The payload was planted in Python start-up files, so it ran whenever Python started in an affected environment, even if LiteLLM was never imported. Aikido describes an infostealer plus a remote access trojan that polls attacker infrastructure for code to run. The stealer targeted AWS, Azure and GCP credentials including instance metadata, SSH and GPG keys, shell history, browser data, Kubernetes service account tokens, Vault tokens, environment variables, auth files for AI coding tools such as Claude Code, Cursor, Gemini CLI and Codex CLI, and the provider keys stored in the LiteLLM Proxy database. Wiz says data was encrypted and sent to the lookalike domain www.litellm[.]dev. As quoted by Help Net Security, LiteLLM said: "All previous versions are safe, and the impacted releases have been pulled." Wiz added: "The most recent safe version is v1.106.2."

Timeline

DateEvent
24 March 2026First LiteLLM supply chain attack: malicious 1.82.7 and 1.82.8 published with a leaked PyPI token (separate incident).
2 September 2026CISA adds the LiteLLM MCP authentication bypass, CVE-2026-59822, to its Known Exploited Vulnerabilities catalogue (separate incident).
19 October 2026Release permissions widened so two engineers can publish bugfix backports, according to LiteLLM (described as "Monday").
21 October 202608:56 to 10:46 UTC: attacker pushes commits to the lookalike Berri-AI/litellm repository (Wiz).
21 October 202610:08 UTC: branch protection override with a developer credential; 11:11 UTC second credential approves production release.
21 October 202611:39 UTC: seven malicious 1.107.0 wheels reach PyPI; 11:46 UTC Docker images published; 11:53 UTC third credential deletes workflow records.
21 October 2026About 13:20 to 13:23 UTC PyPI quarantines and removes the release; Docker Hub tags cleaned by 14:20 UTC. LiteLLM publishes its incident post.
21 October 202616:20 to 17:01 UTC: stolen AWS instance role credentials tested against Amazon Bedrock and model listing calls (Wiz).
23 October 2026LiteLLM adds a note that some earlier assumptions in its post were inaccurate or incomplete.

How it happened: the identity attack path

  1. Tokens leak into logs. GitHub CLI OAuth tokens used by engineers ended up in response bodies that an internal LiteLLM deployment logged, according to LiteLLM's investigation.
  2. Exposed deployment gives access. That deployment was reachable from the internet. LiteLLM found an unexpected user account on it and believes the attacker read the tokens from its logs; the initial way in is unconfirmed.
  3. Owner token disables protection. Using a token belonging to one of two GitHub organisation owners, the attacker overrode branch protection and force-pushed a workflow change pointing releases at a lookalike repository.
  4. Second identity approves the release. Another engineer's token satisfied the production approval gate, so the signed pipeline built and published poisoned wheels and images. Trusted publishing and signing worked as designed for the wrong code.
  5. Third identity covers tracks. A third engineer's token deleted workflow records, reducing the evidence available.
  6. Payload harvests machine credentials. On every Python start the stealer collected cloud, Kubernetes, Vault, AI tool and LLM provider credentials. Stolen AWS role credentials were then tested against Bedrock, which Wiz links to LLMjacking.

Impact

  • Confirmed: A malicious release was distributed through PyPI and Docker Hub and GHCR tags for roughly two to two and a half hours. Wiz counted about 49,000 PyPI downloads, a figure that includes automated pulls.
  • Confirmed downstream misuse: Wiz observed AWS instance role credentials taken from compromised CI runners and Kubernetes nodes being tested against Amazon Bedrock across multiple environments, with IAM enumeration in a few of them.
  • Not disclosed: How many environments exfiltrated data, and whether any LLM provider keys taken from LiteLLM Proxy databases were used. LiteLLM says enterprise customers on its private registry were not affected.
  • Potential: Any secret reachable from a host that ran 1.107.0 with outbound internet access should be treated as stolen, LiteLLM advises, because the absence of the exfiltration marker file proves nothing. The remote access trojan also gave the attacker a route to run further code.
  • Wider: Sources disagree slightly on the PyPI window (LiteLLM gives removal by 13:23 UTC, Aikido gives an install window to 12:47 UTC). Use the wider LiteLLM window when assessing exposure.

What this means for NHI and AI agent security

This belongs on an NHI list because every step ran on tokens rather than passwords. Three OAuth tokens issued to developer tooling were enough to pass a branch protection rule, a human approval gate and a trusted publishing pipeline. Those controls assume the identity at the gate is the person it belongs to. Once a token can act as an organisation owner, release integrity is only as strong as the place that token is stored, and here that place was a log file. The AI angle runs both ways: an AI coding tool put a token in a response body, an AI gateway logged it, and the payload then went after AI tool auth files and LLM provider keys, with stolen cloud credentials tried against Bedrock.

The general lesson is that logs, traces and prompt histories from AI tooling are secret stores whether anyone planned them that way or not. Request and response logging in an AI Gateway and LLM Gateway Guide sense needs redaction, and developer tokens with release rights need the controls set out in our CI/CD Pipeline Identity Security Guide. Compare the earlier March LiteLLM breach: fixing the publishing token did not help when the attacker came back through the people allowed to approve releases. A similar pattern of stolen developer and registry tokens appears in the Tensorlake npm worm and the OVHcloud CLI compromise from the same month.

Recommendations

  • Treat exposed hosts as compromised. If 1.107.0 ran anywhere with outbound internet access, rebuild the environment and rotate every secret it could reach, including instance roles, LLM provider keys, LiteLLM master and virtual keys and database passwords. See our Supply Chain Compromise Response Playbook.
  • Redact tokens from AI gateway and tool logs. Do not log full request or response bodies from AI tools by default, scan existing logs for OAuth tokens and keys, and restrict who can read them. See our AI Gateway and LLM Gateway Guide.
  • Shrink developer token scope and lifetime. GitHub CLI tokens should not carry organisation owner rights day to day; use separate, short-lived elevation for administration and revoke tokens on any suspicion. See our OAuth Token Theft Incident Playbook.
  • Make release approval resistant to stolen tokens. Require approvals from hardware-backed sessions, alert on branch protection overrides and workflow record deletion, and do not widen release rights temporarily without an expiry. See our CI/CD Pipeline Identity Security Guide.
  • Keep internal AI deployments off the internet. Put internal gateways behind private networking and strong authentication, patch promptly and review unexpected accounts.
  • Pin and verify dependencies. Pin LiteLLM to a known good version such as 1.106.2, avoid floating latest image tags and check the source repository a signed artifact was built from. See our Software Supply Chain Security Guide.
  • Watch AI model usage on cloud roles. Alert on Bedrock or model listing calls from roles that never use them, and on identity calls from outside your infrastructure. See our LLMjacking Guide.

Frequently asked questions

Which LiteLLM version was compromised in October 2026?

Version 1.107.0 on PyPI, and the Docker Hub and GHCR tags v1.107.0, latest, main-latest and main-stable during the window on 21 October 2026. LiteLLM says earlier versions are safe and recommends pinning 1.106.2 or another version published before the incident.

Is this the same as the March 2026 LiteLLM breach?

No. In March, versions 1.82.7 and 1.82.8 were uploaded with a PyPI token leaked through a compromised Trivy scanner. In October the attacker used three engineers' GitHub CLI OAuth tokens to push a release through LiteLLM's own hardened pipeline. The September MCP authentication bypass was a separate vulnerability, not a supply chain attack.

What should I do if I installed LiteLLM 1.107.0?

Do not start Python in the affected environment to clean up, because that reruns the malware. Remove the malicious start-up files from the shell, stop the background process, rebuild from scratch on a pinned clean version and rotate every credential the host could reach. Then review cloud logs for identity and Bedrock calls from unfamiliar addresses.

LiteLLM PyPI Package Breach (March 2026) · LiteLLM MCP Auth Bypass 2026 · OpenCode Console Breach 2026 · CI/CD Pipeline Identity Security Guide · LLMjacking Guide

How NHI Mgmt Group can help

LiteLLM was hit twice in one year through two different machine identities, which shows that fixing one token rarely fixes the programme. We help teams find where developer and pipeline tokens live, including in logs, cut their scope and lifetime, and build a rotation plan for AI gateway and provider keys before the next incident. See our NHI and AI agent security training.

References

PDF for subscribers

Take this breach analysis with you

The full analysis in one PDF: timeline, attack path, impact and recommendations, with links to the sources and related pages. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 26 October 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.