Deep Seek Harness Flaw Let AI Agents Disable Their Own File Sandbox Without Approval

Summary: A flaw in DeepSeek Harness, DeepSeek's open-source tool for running AI coding agents locally, let a sandboxed agent disable its own sandbox with a single shell command that called the tool's local web interface and set the session to a mode called danger-full-access. Tracked as CVE-2026-82533 and rated 9.4 out of 10 by VulnCheck, it worked on a default installation until DeepSeek fixed the tool on August 27, and versions 0.1.1-rc.2 and earlier are affected.
Bottom line: The interface had no authentication and its access check read only the client-supplied Host header, so a caller who reached it could also download every stored conversation without a key. The record names 0.1.2-alpha.1 as the fixed version, but that release was never published to npm, where the project's own instructions send users - and the fix adds a token check to the interface while leaving the sandbox's unconfined reads and network access unchanged.

Read the original at The Hacker News ↗

Topics: Vulnerabilities & Patching
Mentions: CVE-2026-82533
Related: ⚡ Weekly Recap: Rogue AI Agents, WeChat Worm, Paper Cut Attacks, AI Espionage, and Rootkits Cyber Risks ·
Attackers Exploit UnrealIRCd Flaw That Enables Server-Level Command Injection Cyber Risks · Nov 14, 2026
U.S. CISA adds Check Point and UnrealIRCd flaws to its Known Exploited Vulnerabilities catalog Cyber Risks · Nov 14, 2026