top of page
ALL POSTS
AutoJack Attack: Web Page Hijacking Enables AI Agent Host Code Execution
Key Findings Microsoft researchers discovered AutoJack, an exploit chain in AutoGen Studio that allows a single web page to execute arbitrary code on a developer's machine The vulnerability exists in the Model Context Protocol (MCP) WebSocket handler and requires no authentication, credentials, or user interaction beyond the agent loading a malicious URL Vulnerable pre-release builds 0.4.3.dev1 and 0.4.3.dev2 were shipped to PyPI, though the stable release 0.4.2.2 is unaffect
Jun 203 min read
Marimo RCE Vulnerability CVE-2026-39987 Under Active Exploitation Since Disclosure
Key Findings Critical RCE vulnerability CVE-2026-39987 in Marimo (CVSS 9.3) exploited within 9 hours 41 minutes of disclosure Unauthenticated attackers can obtain full interactive shell access on exposed instances through /terminal/ws WebSocket endpoint Affects all Marimo versions up to 0.20.4; patched in version 0.23.0 Unknown threat actor built working exploit from advisory alone, with no public PoC available Attacker conducted credential theft operation and reconnaissance,
Apr 102 min read
Flaw in ClawJacked Lets Malicious Sites Hijack Local OpenClaw AI Agents through WebSocket
Key Findings OpenClaw has fixed a high-severity security issue that could have allowed a malicious website to connect to a locally running AI agent and take over control. The flaw, dubbed "ClawJacked" by Oasis Security, enables a malicious website to silently open a WebSocket connection to the local OpenClaw gateway and brute-force the password. Upon successful authentication, the malicious script can register as a trusted device, which is automatically approved by the gatewa
Mar 12 min read
bottom of page
