🛡️Anthropic Lets Claude Run Unattended, Caps Blast Radius With Containment
Your agents need containment, not just permission prompts
TL;DR
Anthropic says the fix for risky AI agents is containment: sandboxes, VMs, and egress controls, not endless permission prompts. Its Claude Opus 4.7 holds attack success near 0.1% on single attempts. If you're deploying agents, this is the blueprint to copy.
Anthropic's engineering team says the way to ship agents safely is containment, not asking a human to click approve every turn. Their framework: risk equals the chance of failure times the potential damage, and they've pushed the first term down with better safeguards and training. The second term, the blast radius, gets capped with sandboxes, virtual machines, and egress controls. If you run Claude Code, claude.ai, or Claude Cowork, this is your architecture doc. Each product has a different audience, so each needs a different containment setup. Claude Code's reference devcontainer exists so the agent can run unattended, and Claude Code auto mode is meant to cut approval fatigue. Defenses should overlap across the environment, the model, and the external content the agent can reach. The numbers: Claude Opus 4.7 holds attack success to roughly 0.1% on single attempts, rising to around 5-6% after 100 adaptive attempts. Claude Code auto mode catches roughly 83% of overeager behaviors before they execute. Gray Swan's Agent Red Teaming benchmark tests prompt injection susceptibility. The catch: the ephemeral container shrinks blast radius to almost nothing, but it also caps what Claude can actually do.
Key Points
Claude Opus 4.7 holds attack success to roughly 0.1% on single attempts, rising to around 5-6% after 100 adaptive attempts, according to Anthropic's reported figures for prompt-injection style attacks.
Keep reading — it's free
Enter your email to keep reading — plus the best of AI & tech, daily. Free, forever.
Already a member? Sign in
Comments
Be the first to comment