An engineer developed a tool called agentblast to audit the security risks associated with LLM agents, particularly those using Claude Code and Codex. The audit of 153 local sessions revealed significant security vulnerabilities, including 247 SSH root calls, plaintext API keys, and an empty permissions deny list. The tool also identified issues with auto-trusted project roots and sensitive data exposure in logs, highlighting the potential for unintended system mutations despite read-only inspection commands. AI
IMPACT Highlights critical security risks in LLM agent usage, prompting developers to implement better safeguards and auditing practices.
RANK_REASON The item describes a new tool developed by an individual to audit LLM agent security.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
