Do Androids Dream of Breaking the Game? Systematically Auditing AI Agent Benchmarks with BenchJack

Agent benchmarks have become the de facto measure of frontier AI competence, guiding model selection, investment, and deployment. However, reward hacking, where agents maximize a score without performing the intended task, emerges spontaneously in frontier models without overfitting. We argue tha...

Command Injection

Overview @wonderwhy-er/desktop-commander is a MCP server for terminal operations and file editing Affected versions of this package are vulnerable to Command Injection via the extractBaseCommand function. An attacker can execute arbitrary operating system commands by supplying crafted input that ...

Linux kernel race konditions

Race conditions on TIOCSETLD during read/write operation on same terminal can cause system to crash and potentially may lead to privilege escalation...