Is this vulnerability remotely exploitable?

It can be indirectly exploitable from remote locations: an attacker who controls or modifies a website or online content that the developer visits can embed malicious instructions that the AI agent misinterprets as helpful guidance, potentially leading to local command execution.

What level of access does an attacker need?

The attacker does not need direct access to the victim's machine but must be able to influence content consumed by the AI assistant (for example, by hosting a malicious documentation page, compromising an existing site, or injecting content into pages the user is likely to visit while using the editor).

What is the safest general fix pattern for this type of issue?

The safest pattern is to treat AI-generated outputs as untrusted data, enforce strong mediation between AI decisions and execution, and require explicit, informed user approval for any security-sensitive action. Execution components should enforce tight whitelisting, argument validation, and sandboxing so that even if the AI proposes harmful commands, they cannot run automatically.

Severity:

HIGH RISK

(8.8)

Indirect Prompt Injection Leading to OS Command Execution (CWE-78) in Cursor AI Code Editor

CVE-2026-31854

Published:

Mar 11, 2026

Status:

Analyzed

Summary

Affected Context

Exploit Preconditions

Why This Happens

Potential Impact

General Mitigation Guidance

How Teams Detect Issues Like This

Frequently Asked Questions

Source

NDV

Recent Vulnerabilities

CWE-918

HIGH RISK

HIGH

(8.6)

Server-Side Request Forgery (CWE-918) in OpenClaw Zalo Plugin sendPhoto

Unauthorized Access to Internal Resources via SSRF

May 6, 2026

CWE-863

HIGH RISK

HIGH

(8.8)

Authorization Bypass (CWE-863) in OpenClaw Matrix Room Control Commands

Privilege Escalation via Bypassed Access Control

May 6, 2026

Vulnerability Type

Category:

AI-Assisted Command Injection

CWE:

CWE-78

Impact:

Unintended OS Command Execution via AI Agent

Severity:

HIGH RISK

(8.8)

CVSS Version:

3.1

Vulnerability Type

Category:

AI-Assisted Command Injection

CWE:

CWE-78

Impact:

Unintended OS Command Execution via AI Agent

Severity:

HIGH RISK

(8.8)

CVSS Version:

3.1

Catch This CVE While Code Is Being Written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Get Started

Ship clean & secure code faster

Start Free Trial

Get Pentest Report

Ship clean & secure code faster

Start Free Trial

Get Pentest Report

Ship clean & secure code faster

Start Free Trial

Get Pentest Report

CodeAnt AI is an offensive and defensive security platform.

Defensive + offensive security.

From first line of code to a pentest
that knows every line of it.

SOC 2 Type II certified | CodeAnt AI is independently audited and compliant with SOC 2 Type II security standards

Ask AI for summary of CodeAnt

Defensive + offensive security.

From first line of code to a pentest
that knows every line of it.

Ask AI for summary of CodeAnt

Defensive + offensive security.

From first line of code to a pentest
that knows every line of it.

Ask AI for summary of CodeAnt

Indirect Prompt Injection Leading to OS Command Execution (CWE-78) in Cursor AI Code Editor

Summary

Affected Context

Exploit Preconditions

Why This Happens

Potential Impact

General Mitigation Guidance

How Teams Detect Issues Like This

Frequently Asked Questions

Source

Recent Vulnerabilities

HIGH RISK

HIGH

(8.6)

Server-Side Request Forgery (CWE-918) in OpenClaw Zalo Plugin sendPhoto

HIGH RISK

HIGH

(8.8)

Authorization Bypass (CWE-863) in OpenClaw Matrix Room Control Commands

Vulnerability Type

Vulnerability Type

Vulnerability Type

Catch This CVE While Code Is Being Written

Ship clean & secure code faster

Ship clean & secure code faster

Ship clean & secure code faster

Product

Compare

Company

Developers

Legal

Product

Compare

Company

Developers

Legal

Product

Compare

Company

Developers

Legal