Claude Code Desktop flaw exposes developer secrets

A new study by Noma Security found a vulnerability in Anthropic’s Claude Code Desktop app that allows attackers to run malicious JavaScript inside the app’s preview pane, giving them access to the host computer’s files and cloud credentials.

The flaw, called SneakPeek, uses a size threshold that changes how the app renders HTML.

An attacker could put a malicious file in a public repository and wait for a developer to ask Claude Code to open it. When the preview pane renders the file, the embedded script can read local files such as AWS keys, GitHub secrets and SSH keys and send them to a remote server.

The weakness has been given a high‑severity CVSS score of 8.7. Anthropic has worked with Noma to fix the problem and is preparing a GitHub Security Advisory while requesting a CVE ID that will be released soon.

Noma’s researchers noted that the first patch blocked direct file reads from the main page, but the attackers moved the read code into a Web Worker, a background thread that the patch did not cover. The team then issued a deeper architectural fix that fully resolved the issue.

The incident shows that security controls designed for the AI model do not automatically protect all parts of the workflow. Experts advise keeping AI tools up to date, avoiding long‑lived credentials on developers’ machines, using scoped short‑lived tokens, and monitoring runtime activity to detect unauthorized file reads and outbound connections.