Type: url
Original: https://zenn.dev/team_zenn/articles/ai-agent-security
Status: generated
Summary
Based on a January 2026 conference talk, this post surveys the security risks of working with AI coding agents — credential exposure, vulnerable or malicious dependencies, prompt injection, and insecure generated code — and argues from a combinatorial estimate that exhaustive output checking by model vendors is not technically possible, so verification falls to the user. It then sets out guardrails against each risk: keeping secrets out of plaintext local storage, cooldown settings and scanners for dependencies, curating permissions.allow/deny so that automatic approval can be used without habituating reviewers, and three sandboxing options — Dev Containers, Claude Code’s built-in sandbox, and cloud IDEs — compared on effort, stability, IDE choice and external-tool integration. The author concludes that no guardrail set removes the need for human supervision, and that review intensity should be scaled to how mission-critical the code is.