Claude Code Safety Fails: Attack Succeeds 80% and Blocks Cleanup
Simon Willison · rss · 2026-08-28
Simon Willison reports that security researcher Johann Rehberger discovered a high-risk attack against Claude Code's "Auto Mode" with an 80% success rate. The attack bypasses detection by tricking the agent into downloading a zip archive and executing code via base64 imports. Critically, the safety mechanism has a logical flaw: it allows the execution of malicious code in some cases but blocks Claude's attempt to terminate the process, turning the safety feature into an obstacle. The conclusion is that unattended agents must run in containers, VMs, or sandboxes with restricted network egress.
More from coding & agent
- Google DeepMind uses Teamwork multi-agent framework for math breakthroughs — algo_diver · 2026-08-28
- Oxlint Config v2 Released to Keep AI Agents on Track — cnakazawa · 2026-08-28
- Overengineering pitfalls: Misusing Docker and excessive constraints — burny_tech · 2026-08-28
- Photographer tests ChatGPT Desktop for photo editing: achieves perfect results, saves $1k/month — Ice2jc · 2026-08-28
- Video: Why teams are taking back control of their AI coding stack — labeveryday · 2026-08-28
- Relay Harness Unifies 204 Models into a Single Coding Agent with Prepaid Billing — realmeetjames · 2026-08-28