# LiteLLM 1.107.0: check where it ran

Canonical: https://www.oktsec.com/radar/litellm-1-107-0
Publisher: Oktsec
Language: en
Published: 2026-10-23T20:34:00Z
Updated: 2026-11-15T13:21:59Z
Record reviewed: 2026-11-15T13:21:59Z
Latest evidence check: 2026-11-15T13:21:59Z
Disclosed: 2026-10-21

Malicious LiteLLM 1.107.0 artifacts require credential rotation and clean rebuilds. Newer releases use keyless cosign; earlier images retain key signatures, including compromised images signed with the old key.

## Evidence and response priority

- Category: Supply chain
- Oktsec priority: Act now
- Rationale: A malicious package was distributed. If it executed, treat credentials accessible to that process as potentially exposed.
- Evidence: Attacks confirmed
- Verification: Vendor confirmed
- Fix status: Additional action required
- Exposure requirement: Affected package must execute
- Vendor: LiteLLM
- Products: LiteLLM

## Technical overview

The definitive October 23 investigation traces the release compromise to an administrative key exposed in a public repository in May and never revoked. The attacker obtained an organization owner’s GitHub token from the Agent Platform on October 10. Stolen identities enabled malicious workflow changes and release approvals. This corrects the initial hypothesis about tokens taken from deployment logs.

Affected distributions include PyPI wheels and Docker Hub/GHCR images. The vendor excludes the official source tag and private enterprise registry. Distribution windows describe availability, not the last possible execution of cached copies.

Oktsec recommends preserving evidence, revoking reachable credentials and rebuilding from a clean environment. Do not start Python or pip in an affected environment: startup files can rerun the payload. A clean replacement does not invalidate stolen credentials.

LiteLLM’s November 15 update changes signature verification from v1.111.0-dev.3 onward: release images and Helm charts use keyless cosign, with a certificate identifying its GitHub release workflow and a Rekor transparency record. Verify the exact workflow identity and issuer listed in the vendor instructions. The command using a public key does not apply to these releases. This boundary does not recommend a development release for production.

Earlier images retain their key signatures. The replacement public-key SHA-256 fingerprint published after the October 23 rotation is e3fe3e08c44b16407fe0b37c43b4f8b603b3a0f4ee74c29e647bc022d762c2c0. Compromised v1.107.0 images also had valid signatures from the previous key. Check digest, publication history and signature provenance together; signature validity alone does not establish safety.

## Recommended response

Identify affected executions, replace accessible credentials and rebuild from a clean environment.

Relevant to: Teams that obtained affected PyPI or container artifacts, including cached copies executed after registry withdrawal.

## Affected systems and fixes

### PyPI wheels

Affected: litellm 1.107.0, published October 21 between 11:39:51 and 13:23 UTC

Fixed release / action: Rebuild using a verified clean release; initial vendor recovery guidance names 1.106.2

### Docker Hub images

Affected: v1.107.0, latest, main-latest and main-stable: compromised artifacts served October 21, 11:46:56–14:20 UTC

Fixed release / action: Use a verified clean image digest

### GHCR images

Affected: Affected tags of litellm, litellm-database and litellm-non_root: October 21, 11:46:56 UTC until removal by 13:51 UTC

Fixed release / action: Use a verified clean image digest

Confirm the running version against the vendor advisory, including any distribution backports.

## Supporting evidence

Snapshot checked: 2026-11-19T21:12:22.272Z. These dates do not change the incident’s editorial review date.

### OSV / OpenSSF package evidence

PyPI: litellm 1.107.0. Query status: available; checked 2026-11-19T21:12:22.272Z; last successful check 2026-11-19T21:12:22.272Z.

- [MAL-2026-2144](https://osv.dev/vulnerability/MAL-2026-2144): Malicious code in litellm (PyPI). Record modified 2026-10-29T14:30:33.719516315Z. Active database record. OpenSSF provenance: verified. [Original OpenSSF record](https://github.com/ossf/malicious-packages/blob/main/osv/malicious/pypi/litellm/MAL-2026-2144.json)

A database match is not a scan of your environment. OSV and an OpenSSF record it imports are one underlying source, not independent confirmations.

## Response checklist

1. **Find where it ran** — Check package records, image digests, cached artifacts and CI deployments against the vendor’s PyPI, Docker Hub and GHCR windows. Cached copies may execute after removal from a registry; record when and where the affected artifact ran.
2. **Replace reachable credentials** — Revoke and replace credentials reachable from an affected environment, including model provider keys, IAM credentials, database passwords and proxy keys. Review their use from outside your infrastructure and preserve relevant audit records.
3. **Rebuild from a clean environment** — Preserve evidence and rebuild from a clean environment. Do not start Python or pip in an affected environment: startup files can rerun the payload. Verify the replacement digest and use the signature method documented for its release; a valid signature alone does not exclude compromised images.
4. **Use the verification method for the selected release** — From v1.111.0-dev.3, LiteLLM documents keyless cosign for release images and Helm charts. Match the exact certificate identity and GitHub OIDC issuer in the vendor’s verification instructions; do not accept an arbitrary signer. Earlier releases still require the appropriate public key. This signing boundary is not a recommendation to deploy a development release or proof of host cleanup.

## Relevance to AI environments

Model provider keys and agent credentials stored in the same environment belong in the investigation.

## Sources

- [October 21 supply chain incident](https://docs.litellm.ai/blog/security-incident-oct-21-2026) — LiteLLM; Vendor notice; checked 2026-11-15T13:21:59Z
- [Definitive investigation: the attack chain behind release 1.107.0](https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain) — LiteLLM; Vendor notice; checked 2026-11-15T13:21:59Z

## Editorial record

- 2026-10-23T20:34:00Z: Published. Malicious distribution confirmed by the vendor; customer impact requires investigation.
- 2026-10-23T21:03:38Z: Added the response priority rationale and original disclosure date.
- 2026-10-23T21:15:06Z: Clarified the vendor warning against running Python or pip in an affected environment, and included this incident in AI and cloud credential coverage.
- 2026-10-24T13:11:44Z: Added the vendor’s definitive October 23 investigation, which corrects the initial access hypothesis; distinguished registry exposure windows and cached artifacts. Response priority remains Act now.
- 2026-10-24T13:27:53Z: Edited earlier editorial notes for readers; original wording remains in version history.
- 2026-10-31T12:46:55Z: Added the October 30 signing-key clarification: compromised images could pass old-key signature verification. Included the replacement public-key fingerprint and digest checks; credential rotation remains necessary.
- 2026-11-15T13:21:59Z: Added the November 15 move to keyless cosign for release images and Helm charts, with workflow identity and issuer checks; retained separate verification for older signed images and incident recovery actions.

## Related research

- [security-recap-october-22-23-2026](https://www.oktsec.com/research/security-recap-october-22-23-2026)

## Use this advisory in Radar

Use the Advisory view to understand the issue and its evidence, Affected systems to compare versions, and Response plan to record applicability, an owner and action progress. Plans are saved only in the current browser and can be downloaded for sharing; they are not a shared team account or an automated scan. Completing a plan does not validate remediation. Talk to Oktsec opens a technical consultation with the advisory and selected service preserved. Scope, fees and access are agreed before any assessment.

[Open advisory](https://www.oktsec.com/radar/litellm-1-107-0) · [Build a response plan](https://www.oktsec.com/radar/litellm-1-107-0#review-workspace)

- [Exposure check consultation](https://www.oktsec.com/contact?source=radar&intent=assessment&item=litellm-1-107-0&service=exposure_check)
- [AI agent review consultation](https://www.oktsec.com/contact?source=radar&intent=assessment&item=litellm-1-107-0&service=agent_review)

## Scope and corrections

Radar is a selected editorial collection, not a complete vulnerability database or a scan of your environment. Priorities are Oktsec analysis. Source confirmation and checklist completion do not establish whether your systems are compromised or secure.

[Sources and methodology](https://www.oktsec.com/radar/about). Report corrections with supporting sources to info@oktsec.com. Do not send secrets or private incident data.

