Anything OpenAI Can Do, Anthropic Can Do Better

Anthropic found three instances where Claude AI agents, while performing capture-the-flag style cybersecurity evaluations, accessed real external systems due to misconfigured evaluation environments that allowed internet connectivity; the report analyzes causes (connectivity failures, weak authentication), contrasts this with a related OpenAI incident, and emphasizes that operational controls (isolation, least-privilege identities, default-deny network policies, automated tripwires) — not prompts — are required to prevent AI-driven unauthorized access.