AI Privacy Incident•October 21, 2026

LiteLLM 1.107.0 release pipeline compromise

Vendor: BerriAI
Product: LiteLLM
Severity: critical
Status: ongoing
Users affected: undisclosed; no install count published for the affected release

Summary

On October 21, 2026, LiteLLM, the open source gateway that routes requests to hosted language model providers, disclosed that version 1.107.0 published to PyPI, Docker Hub, and GHCR carried a credential stealer and a remote code loader. LiteLLM said the packages were live for roughly two hours before removal. The project told users to assume exfiltration on any affected host. Its investigation into the initial access is ongoing.

What happened

  • LiteLLM stated that the malicious PyPI release was available from 11:39 to about 13:23 UTC, and that several Docker and GHCR tags pointed to compromised images until as late as 14:20 UTC.
  • The project said an attacker used developer credentials to push a workflow change that built the release from a lookalike repository, Berri-AI/litellm.
  • Datadog Security Labs reported that the package added startup files that run whenever Python launches, even if LiteLLM is never imported.
  • Datadog said the stealer targeted cloud, developer, AI tool, browser, and LiteLLM proxy credentials, and sent encrypted results to a lookalike domain. It observed no second-stage payload.

Timeline

  • 2026-10-21 - Version 1.107.0 appears on PyPI at 11:39 UTC and in container registries shortly after.
  • 2026-10-21 - LiteLLM restores its release workflow and removes or repoints the affected artifacts.
  • 2026-10-21 - LiteLLM publishes its incident notice.
  • 2026-10-23 - LiteLLM updates the notice and states that some earlier assumptions were inaccurate or incomplete.

What the vendor has confirmed

LiteLLM said it revoked maintainer credentials, reduced release permissions, and restored the workflow. It named 1.106.2 as a safe pin. It told operators to rotate provider keys stored in the proxy database, master and virtual keys, database passwords, and cloud role credentials. Customers on its private enterprise registry were not affected, according to the notice.

What remains unclear

  • The number of hosts that installed the affected release has not been disclosed.
  • LiteLLM said it suspects the developer tokens were taken from logs on an internet-exposed internal deployment, but has not confirmed how that system was accessed.
  • Whether any collected credentials were used downstream has not been reported.

Broader context

This is the second publicly documented compromise of the same gateway package in 2026. Datadog said no link between the two has been established. In both cases the release pipeline, not the gateway code, was where a package holding many provider keys was turned against its users.

Sources

Selvam Sivakumar
Written by

Selvam Sivakumar

Founder, Elephas.app

Selvam Sivakumar is the founder of Elephas and an expert in AI, Mac apps, and productivity tools. He writes about practical ways professionals can use AI to work smarter while keeping their data private.

Related Resources

news

OpenAI Paused Frontier Training Over Misalignment. Here Is What That Means for Your Data.

OpenAI paused frontier AI training over misalignment concerns, a warning for people to limit what confidential work and sensitive data they put into AI tools.

17 min
news

ChatGPT's Computer History: What It Actually Records on Your Mac

OpenAI's ChatGPT Computer History logs clicks, typing, and app switches on your Mac. See what it tracks, where the data goes, and if it's safe.

11 min
news

Does Claude Watermark Text You Wrote Yourself? Mostly Not

Anthropic explained how the Claude watermark works. It attaches only to words Claude chose, so proofreading your own writing leaves almost nothing to detect.

14 min
news

AI Going Rogue: 7 Real Examples From Four Major Labs

AI going rogue is not fiction. Seven real 2026 cases from OpenAI, Anthropic, Meta and Moonshot, what caused each one, and what it means for your own AI tools.

14 min