How Claude Code’s 256KB blind spot put developers’ secrets in the preview pane

Summary

Noma Security found SneakPeek, a flaw in Anthropic's Claude Code Desktop app that could let malicious JavaScript in an HTML file run without further user interaction and expose sensitive host files, such as AWS credentials, GitHub secrets and SSH private keys. The flaw sat in the app's automated preview pane, which renders HTML files generated or modified by Claude. Noma traced it to three gaps: a 256KB file-size threshold that switched files to file:// URLs and bypassed the Content Security Policy, a file:// origin that gave the renderer local filesystem access, and an inactive request filter during static file rendering.

Mitigation

Anthropic fully resolved the issue after a deeper architectural fix, following a bypass of its initial patch that Noma reported. The source states no version number or configuration change.

Related items

• InfoAnthropic’s self-replicating AI warning is alarming deja vuSame vendor · The Verge (AI)
• InfoNvidia, worth $5.5tn, donates $250m to UK for AI safety researchSame vendor · The Guardian Technology
• MediumAI agents aren’t ‘going rogue.’ Frontier labs are skipping security basics.Same vendor · CSO Online
• InfoAnthropic opens Singapore office and signs MOU with the Government of SingaporeSame vendor · Anthropic Research Blog
• InfoRedesigning source control at AnthropicSame vendor · Simon Willison's Weblog