Claude Code Desktop Flaw Let Malicious Previews Steal Developer Credentials
Summary
Noma researchers found that chained flaws in Claude Code Desktop’s preview pane could let a malicious HTML file silently exfiltrate developer credentials. Anthropic patched the issue after researchers bypassed an initial fix.
Key points
- The flaw affected Claude Code Desktop’s automated preview pane for HTML files generated or modified by the platform.
- Noma said missing content security policy, unrestricted origin access, and inactive request filtering could be chained to bypass sandbox protections.
- A malicious repository could include an HTML file padded beyond 256KB; opening it for review or editing would trigger the preview and background credential theft without a click.
- Potentially exposed secrets included AWS and GitHub credentials, API keys, Claude OAuth tokens, and other credentials on the developer’s machine.
- Anthropic issued an initial fix on July 9, but Noma found a bypass in a July 24 retest; a final patch was deployed on Aug. 17, according to Noma.
- Noma researchers advised keeping coding applications updated and avoiding long-lived credentials on developer machines.
Article Details
- Vulnerability Types
- Missing content security policy
- Unrestricted origin access
- Inactive request filtering
- Severity
- High
- Exploitation Status
- unknown
- Exploit Availability
- unknown
- Patch Status
- available
- Workarounds
- Keep coding applications updated through an aggressive, automated patching program.
- Keep long-lived credentials off developer machines.