A cybersecurity researcher says Anthropic has silently patched a vulnerability that would have allowed an attacker to bypass the Claude Code network sandbox, potentially enabling data exfiltration.
Claude Code’s network sandbox funnels all outbound traffic through a local allowlist proxy, silently blocking any connection to unapproved hosts.
According to vulnerability researcher Aonan Guan, two Claude Code network sandbox bypasses were discovered recently. One of them, tracked as CVE-2025-66479 and discovered by a different researcher, was related to the sandbox interpreting a setting to block all outbound traffic as ‘allow everything’.
This issue was fixed with an update released on November 26, 2025.
The second sandbox bypass vulnerability, discovered by Guan, has been described as a SOCKS5 hostname null-byte injection issue.
“The userʼs policy says allow only *.google.com. The attacker sends a hostname like attacker-host.com\x00.google.com. The filter sees the trailing .google.com and approves; the OS truncates at the null byte and dials attacker-host.com,” Guan explained.
According to Guan, the vulnerability was present in the Claude Code network sandbox from October 20, 2025, when the sandbox became generally available, until the release of version 2.1.90 in April, around the time he reported it through Anthropic’s bug bounty program on HackerOne. The AI giant marked the vulnerability report as a duplicate.
The researcher is displeased that Anthropic has not assigned a CVE identifier to this vulnerability and has not mentioned the issue in its release notes.
Moreover, Guan noted that CVE-2025-66479 was assigned to the ‘sandbox-runtime’ library rather than Claude Code itself, and there was no warning to Claude Code users.
“A team running [the vulnerable configuration] in production from October 20 through November 26 had no way to know the sandbox was effectively off, and no notice afterwards that it had ever been off. The CVE shipped against a library most Claude Code users do not know exists by name,” the researcher said.
Guan recently disclosed details of a prompt injection attack method called Comment and Control. The attack worked against popular AI code security and automation tools, including Claude Code Security Review, Gemini CLI Action, and GitHub Copilot Agent.
He and other researchers discovered that AI agents associated with these tools on GitHub Actions could be hijacked using specially crafted GitHub comments, including PR titles, comments, and issue bodies.
In his disclosure of the Claude Code sandbox vulnerability, Guan noted that the bypass would have been particularly useful in combination with a prompt injection attack such as Comment and Control, enabling attackers to exfiltrate data, including environment variables, credentials, tokens, and infrastructure data.
Contacted by SecurityWeek, Anthropic said it appreciates Guan’s work, but its security team had identified and fixed this issue before receiving the researcher’s report.
The AI giant clarified that the fix was included in a public commit to the ‘sandbox-runtime’ repository on March 27 and was shipped in Claude Code 2.1.88 on March 31, before Guan submitted his report via HackerOne on April 3.
Related: ‘Claw Chain’ OpenClaw Flaws Allow Sandbox Escape, Backdoor Delivery
Related: OpenAI Hit by TanStack Supply Chain Attack
Related: Mythos Proves Potent in Vulnerability Discovery, Less Convincing Elsewhere
