CVE-2026-59820: LiteLLM Zip Slip in Skills upload

Q: Is CVE-2026-59820 actively exploited?

No confirmed active exploitation of CVE-2026-59820 has been reported, but organizations should still patch proactively.

Q: How to fix CVE-2026-59820?

Upgrade to LiteLLM 1.83.7-stable or later immediately — the fix is in the referenced commit and release. Until patched, audit API key configurations and remove /v1/skills, anthropic_routes, and llm_api_routes from allowed_routes for any key or user that doesn't explicitly need Skills functionality, minimizing the population that can trigger the vulnerable code path. Run the LiteLLM proxy process with the least filesystem privilege possible (no write access to config, code, or other sensitive directories beyond the intended skills staging path). For detection, monitor skill upload requests for ZIP entries containing '../' or absolute paths, and watch the filesystem around the extraction/staging directory for writes landing outside expected boundaries.

Q: What systems are affected by CVE-2026-59820?

This vulnerability affects the following AI/ML architecture patterns: LLM gateway/proxy, agent frameworks, multi-tenant LLM API access.

Q: What is the CVSS score for CVE-2026-59820?

CVE-2026-59820 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.59%.

CISO Take

LiteLLM, a widely deployed AI gateway/proxy that fronts LLM API calls, shipped a Skills archive extraction routine that didn't validate file paths inside uploaded ZIP files, letting an authenticated user or API key with access to skills-related routes write files outside the intended staging directory via classic zip-slip path traversal. This isn't a pre-auth, internet-wide smash-and-grab — it requires a valid key whose allowed_routes include /v1/skills, anthropic_routes, or llm_api_routes — and the EPSS score of 0.00323 confirms real-world exploitation is currently very unlikely, with no public PoC or Nuclei template observed and no CISA KEV listing. That said, LiteLLM sits at the center of many organizations' AI infrastructure as the single gateway multiplexing calls across models and tenants, so an authenticated-but-lower-trust key being able to write files on the proxy host is a meaningful escalation path, particularly in multi-tenant or shared-key deployments. Patch to 1.83.7-stable now; in the interim, audit which API keys have skills/anthropic_routes/llm_api_routes access and strip that scope from any key that doesn't strictly need it.

Sources: NVD GitHub Advisory EPSS CISA KEV ATLAS

What is the risk?

No official CVSS score is published, but the underlying flaw (CWE-22, path traversal / zip-slip during archive extraction) has a moderate-to-high theoretical severity because it enables arbitrary file writes on the LiteLLM proxy host. Real-world risk is currently tempered by three factors: exploitation requires an authenticated user or API key already scoped to skills-capable routes (not unauthenticated internet exposure), EPSS sits at 0.00323 (a low near-term exploitation probability despite the percentile framing), and there is no public exploit code, Nuclei template, or CISA KEV entry. Risk rises sharply in environments that issue skills-scoped API keys broadly (e.g., to less-trusted internal teams, partners, or automated agents) or that run the LiteLLM proxy process with write access to sensitive paths.

How does the attack unfold?

Initial Access

Attacker obtains or already holds a valid LiteLLM API key/account whose allowed_routes include /v1/skills, anthropic_routes, or llm_api_routes.

AML.T0012

Malicious Upload

Attacker crafts a skill ZIP archive with path traversal entries (e.g. '../../') and submits it through the skills upload endpoint.

AML.T0049

Unsanitized Extraction

LiteLLM's archive extraction logic fails to validate resolved paths, writing archive contents outside the intended extraction/staging directory.

Host Impact

Files written outside the sandboxed directory can overwrite configuration or other files reachable by the LiteLLM process, escalating a narrowly-scoped credential toward broader gateway host compromise.

AML.T0079

Initial Access

Attacker obtains or already holds a valid LiteLLM API key/account whose allowed_routes include /v1/skills, anthropic_routes, or llm_api_routes.

AML.T0012

Malicious Upload

Attacker crafts a skill ZIP archive with path traversal entries (e.g. '../../') and submits it through the skills upload endpoint.

AML.T0049

Unsanitized Extraction

LiteLLM's archive extraction logic fails to validate resolved paths, writing archive contents outside the intended extraction/staging directory.

Host Impact

Files written outside the sandboxed directory can overwrite configuration or other files reachable by the LiteLLM process, escalating a narrowly-scoped credential toward broader gateway host compromise.

AML.T0079

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
LiteLLM	pip	< 1.83.7	`1.83.7`
59.7K OpenSSF 5.9 6 dependents Pushed 5d ago 55% patched ~46d to patch Full package profile →

Do you use LiteLLM? You're affected.

How severe is it?

CVSS 3.1

6.5 / 10

EPSS

0.6%

chance of exploitation in 30 days

Higher than 46% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Moderate

What is the attack surface?

AV Network

AC Low

PR Low

UI None

S Unchanged

C None

I High

A None

What should I do?

1 step

Upgrade to LiteLLM 1.83.7-stable or later immediately — the fix is in the referenced commit and release. Until patched, audit API key configurations and remove /v1/skills, anthropic_routes, and llm_api_routes from allowed_routes for any key or user that doesn't explicitly need Skills functionality, minimizing the population that can trigger the vulnerable code path. Run the LiteLLM proxy process with the least filesystem privilege possible (no write access to config, code, or other sensitive directories beyond the intended skills staging path). For detection, monitor skill upload requests for ZIP entries containing '../' or absolute paths, and watch the filesystem around the extraction/staging directory for writes landing outside expected boundaries.

What does CISA's SSVC say?

Decision Track

Exploitation none

Automatable No

Technical Impact total

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Code Execution API Plugin AML.T0012 - Valid Accounts AML.T0049 - Exploit Public-Facing Application AML.T0079 - Stage Capabilities

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.2 - AI system security and vulnerability management

NIST AI RMF

MANAGE-4.1 - AI system risks and benefits are monitored, and post-deployment risk management follows established procedures

OWASP LLM Top 10

LLM03 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2026-59820?

LiteLLM, a widely deployed AI gateway/proxy that fronts LLM API calls, shipped a Skills archive extraction routine that didn't validate file paths inside uploaded ZIP files, letting an authenticated user or API key with access to skills-related routes write files outside the intended staging directory via classic zip-slip path traversal. This isn't a pre-auth, internet-wide smash-and-grab — it requires a valid key whose allowed_routes include /v1/skills, anthropic_routes, or llm_api_routes — and the EPSS score of 0.00323 confirms real-world exploitation is currently very unlikely, with no public PoC or Nuclei template observed and no CISA KEV listing. That said, LiteLLM sits at the center of many organizations' AI infrastructure as the single gateway multiplexing calls across models and tenants, so an authenticated-but-lower-trust key being able to write files on the proxy host is a meaningful escalation path, particularly in multi-tenant or shared-key deployments. Patch to 1.83.7-stable now; in the interim, audit which API keys have skills/anthropic_routes/llm_api_routes access and strip that scope from any key that doesn't strictly need it.

Is CVE-2026-59820 actively exploited?

No confirmed active exploitation of CVE-2026-59820 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-59820?

Upgrade to LiteLLM 1.83.7-stable or later immediately — the fix is in the referenced commit and release. Until patched, audit API key configurations and remove /v1/skills, anthropic_routes, and llm_api_routes from allowed_routes for any key or user that doesn't explicitly need Skills functionality, minimizing the population that can trigger the vulnerable code path. Run the LiteLLM proxy process with the least filesystem privilege possible (no write access to config, code, or other sensitive directories beyond the intended skills staging path). For detection, monitor skill upload requests for ZIP entries containing '../' or absolute paths, and watch the filesystem around the extraction/staging directory for writes landing outside expected boundaries.

What systems are affected by CVE-2026-59820?

This vulnerability affects the following AI/ML architecture patterns: LLM gateway/proxy, agent frameworks, multi-tenant LLM API access.

What is the CVSS score for CVE-2026-59820?

CVE-2026-59820 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.59%.

What is the AI security impact?

Affected AI Architectures

LLM gateway/proxyagent frameworksmulti-tenant LLM API access

MITRE ATLAS Techniques

AML.T0012 Valid Accounts

AML.T0049 Exploit Public-Facing Application

AML.T0079 Stage Capabilities

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.6.2

NIST AI RMF: MANAGE-4.1

OWASP LLM Top 10: LLM03

What are the technical details?

Original Advisory

LiteLLM is a proxy server (AI Gateway) to call LLM APIs in OpenAI (or native) format. Prior to 1.83.7-stable, LiteLLM Skills archive extraction did not sufficiently validate file paths from uploaded skill ZIP archives, allowing an authenticated user with access to LiteLLM LLM API routes or a key whose allowed_routes includes /v1/skills, anthropic_routes, or llm_api_routes to upload a crafted skill archive containing path traversal entries that could be written outside the intended extraction or staging directory. This issue is fixed in version 1.83.7-stable.

Exploitation Scenario

An attacker obtains or is issued a legitimate LiteLLM API key scoped to skills-related routes — for example, a contractor, a lower-trust internal team, or a compromised automation credential with narrow intended permissions. They craft a ZIP archive containing entries with path traversal sequences (e.g., '../../../etc/some-config') and upload it through the /v1/skills endpoint as if registering a normal skill. LiteLLM's extraction logic writes the archive's contents to disk without validating that resolved paths stay within the intended staging directory, so the attacker's traversal entries land outside that boundary — potentially overwriting proxy configuration or other files the LiteLLM process later reads or executes, escalating a narrowly-scoped API credential into broader control over the gateway host.

Weaknesses (CWE)

CWE-22 Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal') Primary CWE-22 Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal') Primary CWE-22 Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal')

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal'): The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted directory.

[Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
[Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.

Source: MITRE CWE corpus.