CYOPS · Cyber Operations

LiteLLM 1.107.0: an unrevoked key, eleven quiet days, then a poisoned release

The maintainers' corrected root cause traces the 21 October compromise to a master key leaked in May and never revoked. The attacker held organisation-owner access for at least eleven days before publishing a credential stealer to PyPI and Docker.

The LiteLLM compromise of 21 October was not a fast smash-and-grab. The maintainers’ canonical root-cause record shows that the attacker entered with a master key committed to a public repository on 27 May, identified for revocation, and never revoked [1]. That key returned an organisation owner’s GitHub CLI token on 10 October. Eleven days later the attacker used it to point the official release workflow at a lookalike repository and publish a credential stealer under the project’s own signing and registry identities [1][5].

LiteLLM is a widely deployed gateway that gives applications one interface to many model providers. Socket and The Register put its PyPI traffic at about 23 million downloads a week; Mend and BleepingComputer give about 92 million a month [4][9][12][14]. A compromised install sits next to the provider keys, cloud identities and database strings the gateway exists to hold [14].

What the data says

All seven 1.107.0 wheels on PyPI were malicious. They were uploaded between 11:39:51 and 11:40:13 UTC and removed at 13:23 [1][2]. Docker Hub tags for the release, and the floating latest, main-latest and main-stable tags, served the poisoned image from 11:46:56 to 14:20 UTC. Equivalent GHCR images were unavailable by 13:51 [2]. Third-party timings differ slightly. Wiz recorded PyPI quarantine at 13:20 and container removal at about 13:53. Mend gives 14:32 for the Docker images [6][9]. The official source tag did not contain the payload; the artefacts were built from an attacker-controlled fork [2][3].

The maintainers have published no download count. Wiz, citing ClickPy data, and StepSecurity both give about 49,000 PyPI downloads. StepSecurity notes the figure includes CI caches and mirrors and excludes container pulls [6][7]. The true exposed population is therefore unknown in both directions.

The canonical record sets out the chain [1]. The leaked key gave API access to an internal alpha agent-orchestration service. On 10 October a single request to that service returned a stored GitHub CLI OAuth token belonging to one of two organisation owners. The token was not scoped down and was not tied to a service account. Within an hour, workflows on throwaway branches posted repository and organisation secrets, including two further maintainers’ tokens, to an external collector; the runs were then deleted. A second round followed on 12 October [1]. Socket reconstructed these runs from Actions metrics, the activity log and environment deployment records that survived deletion. It counts 20 distinct secret names across the two workflows. Socket states that it cannot establish whether 10 October was the start or how the collected credentials were used [5].

On 21 October the attacker disabled branch protection with the owner token, redirected the release workflow to the lookalike organisation, and used a second maintainer’s token to pass the production approval gate. A third token deleted the workflow run records and their logs [1]. The record notes that release approval had been relaxed that week to allow any two engineers to trigger and approve. Previously it required a dedicated account locked to a physical device [1]. StepSecurity dates that change to 19 October and registration of the exfiltration domain to 13 October [7].

The first incident post, published on 21 October, gave a different account. It suspected tokens had been read from logs of an internet-exposed internal deployment and a recently patched vulnerability [2]. On 23 October the maintainers marked that version as inaccurate or incomplete and stated that the first token came from a different service [1][2]. Some early analyses, including a Cloud Security Alliance research note and Help Net Security’s report, reflect the superseded account [10][13].

The payload itself is a startup-hook implant. Microsoft maps it to MITRE ATT&CK T1546.018. A path configuration file and a sitecustomize module installed outside the package directory execute on every Python start, so no application needs to import LiteLLM [3]. A 16-module collector targets cloud, Kubernetes and CI credentials, SSH and GPG keys, browser data, shell histories, and configuration for AI coding assistants. Where a database URL is present it reads provider credentials from the LiteLLM proxy tables [3][4][6]. Output is encrypted with AES-256-CBC under an RSA-4096-wrapped key and posted to a lookalike domain [3][12]. On Unix hosts a detached loader polls a GitHub-hosted pointer hourly and would execute unsigned code from whatever server it names [3][7][8]. The pointer was empty when Microsoft and Socket checked, and the account hosting it was later removed. StepSecurity notes that infected hosts keep polling regardless [3][4][7].

Wiz observed harvested AWS instance-role credentials being validated from external infrastructure about 4.5 hours after exfiltration. In some environments IAM enumeration and discovery followed. Wiz assesses this as consistent with LLMjacking [6]. Help Net Security, summarising Wiz, reports that the checks targeted Amazon Bedrock model access [13].

What it means

No researcher has attributed the activity to a named actor [1][3][6]. This is LiteLLM’s second supply-chain compromise of 2026. In March, malicious 1.82.7 and 1.82.8 builds used a similar startup-file technique and hybrid encryption scheme [11][12]. The Cloud Security Alliance note cites Socket as saying public evidence does not establish a link between the two incidents [10]. The overlap in tradecraft is not attribution.

The decisive control failure was revocation, not detection. Security vendors flagged the wheels within minutes and a user raised a public issue about 20 minutes after release [1][12]. The attacker had already held owner-level access for eleven days. Agent-orchestration services that store developer tokens became the pivot. Those tokens were long-lived, post-two-factor, and broad enough to override branch protection [1].

What to do

  • Determine whether 1.107.0 or the affected floating Docker tags ran anywhere on 21 October, including CI runners, caches, mirrors and lock files. Check by digest, not tag [2][3].
  • Investigate from a known-clean host. Do not run Python or pip in a suspect environment, because any interpreter start re-executes the implant [2][3].
  • Rebuild affected environments and images from verified artefacts pinned to 1.106.2 or an earlier explicit version. Kill the detached loader process. Deleting the package directory alone is insufficient [2][3][9].
  • Assume exfiltration on any host with outbound access. Rotate every reachable secret: model-provider keys, gateway master and virtual keys, database credentials, cloud role credentials, CI tokens, SSH and GPG keys and browser sessions [2][3].
  • Review cloud audit logs for identity checks and model-service enumeration using workload credentials from outside your infrastructure in the hours after install [2][6].
  • Treat agent platforms and gateway logs as credential stores. Keep their admin interfaces off the internet, scan stored content for tokens, and keep organisation-owner credentials off agent-connected workstations [1][10].
  • For your own release pipelines: revoke leaked keys on discovery, require hardware-bound approval for releases, verify build provenance names the expected source repository, and alert on short-lived branches whose workflow runs are deleted [1][5][7].

The cyber operations view

The operator here was patient and audit-aware. They spent eleven days harvesting secrets, used separate maintainer identities to trigger, approve and clean up so that no single account looked anomalous, and erased the run logs within minutes of publication [1][5]. The payload was built to keep working after takedown. It survives package removal and polls for a second stage after the dead drop is gone [3][7]. For defenders, the lesson is that AI infrastructure now concentrates the credentials an intruder wants. The release pipeline that ships it should be treated as tier-zero, and a leaked key is an open incident until it is revoked.

Sources

Every R3KONX article cites its primary material. 14 sources, in order of first citation. Links open the original publication.

  1. LiteLLM v1.107.0 compromise: root cause and attack chain LiteLLM (BerriAI) · 2026-10-23
  2. Contained: Oct 21, 2026 LiteLLM release v1.107.0 compromised LiteLLM (BerriAI) · 2026-10-21
  3. Inside the LiteLLM 1.107.0 supply chain compromise: credential theft through Python startup hooks Microsoft Security · 2026-10-23
  4. LiteLLM PyPI package and Docker images compromised to steal cloud and AI credentials Socket · 2026-10-21
  5. New evidence shows LiteLLM secret-stealing workflows ran 11 days before supply chain attack Socket · 2026-10-23
  6. LiteLLM compromised, steals credentials Wiz · 2026-10-21
  7. litellm 1.107.0 compromised on PyPI: backdoored wheels drop a credential stealer and an hourly remote code loader StepSecurity · 2026-10-21
  8. Compromised LiteLLM 1.107.0 on PyPI installs a credential stealer and backdoor SafeDep · 2026-10-21
  9. LiteLLM 1.107.0 compromised: credential stealer Mend.io · 2026-10-21
  10. LiteLLM 1.107.0: when model-gateway logs become the credential store Cloud Security Alliance · 2026-10-23
  11. LiteLLM supply chain breakdown Upwind · 2026-03-24
  12. LiteLLM release pipeline hacked to distribute info-stealer malware BleepingComputer · 2026-10-22
  13. LiteLLM hit by second supply chain attack this year, users urged to rotate credentials Help Net Security · 2026-10-22
  14. LiteLLM supply chain attack hoovers up passwords and secrets The Register · 2026-10-21

Researched and written by the R3KONX analysis desk from the cited primary material. The maintainers revised their account of initial access on 23 October; this article follows the revised record and notes where earlier reporting relied on the superseded version. Download figures come from third-party analyses of public package statistics; the maintainers have not published exposure numbers. A separate Datadog Security Labs analysis could not be retrieved in full and is not cited. Corrections to event@r3konx.asia.

R3KONX 2027

Work in this domain? So does the programme

4–6 May 2027, World Trade Centre Kuala Lumpur. Curated talks, hands-on training and The BattleGrid.