Arbitrary Command Execution Vulnerability in Claude Agent SDK by Anthropic
CVE-2026-96620

9.2CRITICAL

Key Information:

Vendor

Anthropics

Status
Claude-agent-sdk-python
Vendor
CVE Published:
16 October 2026

What is CVE-2026-96620?

The Claude Agent SDK, a Python SDK for Claude Code, contains a vulnerability that allows an attacker to exploit the resume option passed to the Claude CLI. This flaw manifests when untrusted input is used as the session ID during the resume process. As a result, an attacker can inject the '--mcp-config' flag to set a stdio MCP server, leading to arbitrary command execution at CLI startup beyond the control of the SDK's permission system. This issue has been addressed in version 0.2.121, where appropriate input validation has been implemented to mitigate the risk of this vulnerability.

Affected Version(s)

claude-agent-sdk-python < 0.2.121

References

https://github.com/anthropics/claude-agent-sdk-python/sec...
x_refsource_CONFIRM
https://github.com/anthropics/claude-agent-sdk-python/pul...
x_refsource_MISC
https://github.com/anthropics/claude-agent-sdk-python/com...
x_refsource_MISC

CVSS V4

Score:
9.2
Severity:
CRITICAL
Confidentiality:
High
Integrity:
High
Availability:
High
Attack Vector:
Network
Attack Complexity:
Low
Attack Required:
Physical
Privileges Required:
Undefined
User Interaction:
None

Timeline

  • Vulnerability published

  • Vulnerability Reserved

.