Skip to content
Marc Allington
Marc Allington
← All signals
AI securitybad news for the people this stream is for

Google Threat Intelligence Group says criminal and state-backed actors now use agent-based offensive frameworks to find and exploit vulnerabilities. It found an exposed attacker server running a multi-agent framework called exploit_agent, and describes a group that altered Claude Code settings so the tool fetched a backdoor on launch. Google expects mostly automated attacks within months. Teams should shorten patch times, review who can change coding agent configuration files, and treat public proof-of-concept code with care, as many of those after the recent PHP flaws were AI-generated.