LiteLLM 1.107.0 Supply-Chain Compromise: PyPI Removed and Docker Floating Tags Yanked (October 21, 2026)

The public LiteLLM 1.107.0 release was a supply-chain compromise. All seven PyPI wheels carried a credential stealer and a remote-code loader, and the floating Docker tags on ghcr.io and docker.io were affected; PyPI has removed 1.107.0 and the floating tags are confirmed removed. Here is what executed, what the vendor has confirmed and what it has not, and the practical response for any operator whose environment ran 1.107.0.

Reader ratingRate this article

Loading reader rating…

Privacy-conscious voting uses a signed random browser identifier and a short-lived network hash; the rating tables contain no raw IP-address column.

Reader ratings require JavaScript; the article remains fully readable without it.

Security incident follow-up

LiteLLM 1.107.0: stolen credentials were abused after the poisoned release

This update replaces our October 21 body with the later attack chain, observed post-exploitation activity and a response order designed to avoid re-triggering the malware.

Reporting cutoff: October 22, 2026, 6:59 a.m. EDT / 10:59 UTC.

Act as an incident, not a routine update. If an environment started Python with litellm==1.107.0, or ran an affected Docker image during the published exposure window, LiteLLM says to treat it as compromised. Do not begin recovery by running pip install inside that environment because Python startup executes the malicious files again.[1]
7 wheelsAll PyPI wheels for litellm==1.107.0 were affected.[1]
4 affected tagslatest, main-latest, main-stable and the versioned tag v1.107.0 pointed to compromised container artifacts during the incident window.[1]
~4.5 hoursWiz says it observed stolen AWS role credentials being tested about four and a half hours after exfiltration.[2]

What changed after the first alert

LiteLLM now says the attacker used GitHub CLI credentials tied to three developer accounts to alter, approve and conceal an official release run. The poisoned artifacts were built from a lookalike Berri-AI/litellm repository rather than the legitimate BerriAI/litellm repository, while the official v1.107.0 source tag did not contain the malicious files.[1]

The company says it found an unexpected account in an internet-exposed internal LiteLLM deployment. It believes the attacker reached that deployment, obtained GitHub CLI OAuth tokens from logs, and then used them against the release workflow; it had not confirmed the initial entry path at the reporting cutoff and only suspected a recently patched vulnerability.[1]

Wiz reports that between 16:20 and 17:01 UTC on October 21, the attacker checked harvested AWS instance-role credentials with sts:GetCallerIdentity, then called bedrock:ListFoundationModels and ListModels. Wiz characterizes that sequence as consistent with LLMjacking and says a smaller number of environments also saw IAM and general cloud discovery.[2]

Exact exposure test

You should enter incident response if either of these conditions is true: a Python environment installed litellm==1.107.0 and started Python, or a container executed one of the affected public images during the exposure window. LiteLLM lists the PyPI window as 11:39:51–13:23 UTC and the affected Docker window as 11:46:56–14:20 UTC on October 21.[1]

The affected public container tags were v1.107.0, latest, main-latest and main-stable across the named Docker Hub and GHCR LiteLLM images. LiteLLM says its private enterprise registry was not affected.[1]

Surface Affected artifact Exposure window (UTC, Oct. 21) Status at cutoff
PyPI litellm==1.107.0, all seven wheels 11:39:51–13:23 Removed
Docker Hub v1.107.0, latest, main-latest, main-stable 11:46:56–14:20 Removed or repointed to verified clean artifacts
GHCR The same tags for litellm, litellm-database and litellm-non_root 11:46:56–13:51 Affected tags unavailable by the vendor’s verification

Important: current registry state cannot prove historical safety. A floating tag may point somewhere clean today even if a host pulled the poisoned image during the incident window. Verify the digest actually executed in runtime, deployment, node and registry records.

How the malicious package ran

The poisoned wheels installed cook.pth, cook.py and sitecustomize.py at the site-packages root. The .pth hook imported the stealer whenever Python started; the second startup hook attempted to maintain a detached loader that could fetch and execute another Python stage.[3]

Wiz says the stealer targeted browser cookies and credentials, cloud credentials, AI-provider keys, SSH and GPG material, shell history, process environment variables, and credentials stored in the LiteLLM proxy database. It encrypted collected data and sent it to an attacker-controlled lookalike domain.[2]

The remote-code stage was capable, but Wiz says the dead-drop file was never populated beyond test values during its analysis. Bounded conclusion: that supports “not observed active,” not “incapable” and not proof that every affected host avoided further execution.[2]

Indicators defenders can check

Indicator Why it matters Limit
cook.pth, cook.py, sitecustomize.py in site-packages Files installed by the poisoned wheels Absence does not prove a host is clean
/tmp/.pc_once_72129118c59f5636c0bb Marker written after successful exfiltration Absence does not rule out exfiltration
/tmp/remote_code_loader.lock Lock associated with the detached loader Absence alone does not establish safety
www.litellm[.]dev/healthcheck Attacker-controlled exfiltration endpoint reported by Wiz Practical limit: Network visibility may be incomplete
168.100.8[.]57 and 165.227.201[.]8 Sources Wiz observed validating stolen AWS role credentials Practical limit: An IP match is investigative evidence, not attribution by itself
sts:GetCallerIdentity followed by model-listing calls Sequence Wiz observed in the credential-abuse phase Practical limit: These APIs also have legitimate uses; correlate identity, source and time

Practical caution: LiteLLM explicitly says that finding one of the named files confirms exposure, while not finding one does not prove the environment is clean. Apply that same caution to temporary markers and incomplete logs.[1]

Response order for an affected environment

Practical guidance: the sequence below combines the vendor’s remediation with preserve-before-remediation incident-response practice; it is not a measured guarantee that collection stops immediately.

  1. Contain without starting Python. Restrict outbound access and remove the affected workload from service through the surrounding host, orchestrator or network controls. Preserve short-lived container, process, filesystem, deployment and cloud audit evidence before rebuilding. This sequencing is practical incident-response guidance derived from the startup behavior; it is not a guarantee that collection stops instantly.
  2. Establish artifact history. Record package metadata, image digests, pull times, deployment revisions and which identities or secrets the environment could reach. Do not rely on what a floating tag resolves to now.
  3. Remove the startup files from outside the affected interpreter. LiteLLM directs users to delete cook.pth, cook.py, sitecustomize.py and matching bytecode from every affected site-packages directory, then terminate the detached Python daemon associated with /tmp/remote_code_loader.lock.[1]
  4. Rebuild rather than merely uninstall. LiteLLM recommends rebuilding images and virtual environments from scratch and temporarily pinning open-source deployments to litellm==1.106.2 or an explicit earlier version.[1]
  5. Practical guidance: rotate reachable trust from a clean administrative system. The vendor specifically calls out instance or task IAM roles, provider keys in the LiteLLM proxy database, master and virtual keys, the DATABASE_URL password and secrets exposed through process environments.[1]
  6. Hunt for use after theft. Review CloudTrail around installation time for GetCallerIdentity, then look for use of the role outside expected infrastructure and for the model-listing, IAM-enumeration and discovery activity described by Wiz.[1][2]

Recovery verification

  • Confirm the running package version or image digest on every worker, runner, node and replica—not only in the deployment file.
  • Confirm the three startup files and matching bytecode are absent from every relevant site-packages path.
  • Verify that replacement secrets were issued from a clean system and that old credentials are rejected.
  • Recheck cloud and provider audit logs after rotation for continued use of revoked identities.
  • Keep the incident open if evidence retention, credential scope or historical image identity remains unresolved.

What remains unknown

  • LiteLLM had not confirmed how the attacker first entered the internal deployment; the company described a recently patched vulnerability only as a current suspicion.[1]
  • The reviewed sources do not provide a total victim count or a count of successful exfiltrations.
  • Wiz observed real credential abuse, but reported that the remote-code dead drop was not populated beyond test values during its analysis.[2]
  • No CVE identifier accompanies the malicious-release incident in the cited sources. Do not invent one or conflate the poisoned package with the still-unconfirmed initial-access vulnerability.

Sources

  1. LiteLLM — “Contained: Oct 21, 2026 LiteLLM release v1.107.0 Compromised” (last updated October 21, 2026, 20:30 UTC; accessed October 22, 2026).
  2. Wiz Research — “LiteLLM Compromised, Steals Credentials” (October 21, 2026; accessed October 22, 2026).
  3. BerriAI/litellm issue #47826 — technical disclosure and remediation thread (opened October 21, 2026; accessed October 22, 2026).