[Beginner-oriented] Can Codex and Claude Code be hijacked? Explaining the AGENTS.md attack mechanism and guardrail countermeasures in dialogue form

The Qiita author, in dialogue form and aimed at beginners, explains AGENTS.md/CLAUDE.md prompt injection attacks against Codex and Claude Code and how to configure guardrails. The article cites a research report published on October 26, 2026, stating that an attacker placed a malicious instruction file in the Downloads folder; Codex was induced in over 90% of attempts to add a malicious MCP server and start a session with the sandbox disabled, and Claude Code had API connection settings such as ANTHROPIC_BASE_URL rewritten in about half of attempts. The report also confirmed access using the obtained OAuth token, and the success rates are limited to the researchers' specific validation conditions. The author states that the attack content is based on public research, that configuration examples follow official documentation, and that they were not verified in the author's own environment.