Arbitrary Code Execution in AI Agents: Where the Boundary Breaks
AI agents turn code execution into a security boundary problem when untrusted input can influence commands under real credentials.
Read articleAgentGuard Research
Practical guidance, independent reviews, and clear explanations for teams building AI agents that can take real action.
AI agents turn code execution into a security boundary problem when untrusted input can influence commands under real credentials.
Read articleSafe ChatGPT use starts with data classification, approved accounts, minimal inputs, source checking, and care with connected tools.
Read articleCLI agents need a narrow workspace, explicit command policy, isolated secrets, and tests that inspect actual filesystem and network effects.
Read articleModel security scanning combines artifact inspection with behavioral tests; neither one can cover the release alone.
Read articleUse the OWASP LLM Top 10 as a risk index, then map each relevant item to a deployed path, control, and test.
Read articleEffective AI governance assigns owners and decisions to real systems, then checks whether controls still work after change.
Read articleAI privacy work starts with a traceable data flow: what enters, where it persists, who can retrieve it, and what leaves.
Read articleAudit an AI system by tracing its decisions, data, tools, and real effects rather than reviewing policy documents alone.
Read articleA risk register becomes useful when every important risk maps to an owner, control, test, release decision, and recovery action.
Read articleA practical command-control pattern for coding agents with shell, filesystem, network, and credential access.
Read articleA deployment guide for putting enforceable controls between an agent's intent and its real-world side effects.
Read articleDraw an AI agent threat model that connects attacker influence, trust boundaries, identities, controls, and observable asset changes.
Read article