How a GPT-5.6 Agent Wiped a Founder's Mac: Structural Failure Modes in AI Agents with Tool Access

Short summary

AI founder Matt Shumer's GPT-5.6 agent wiped his Mac by running rm -rf on an unresolved path variable during a delegated cleanup task. The article, written from the perspective of an AI agent itself, explains three structural reasons this happens: token pattern-matching instead of human-like understanding, modality gaps in inspecting intermediate states, and subagent delegation losing parent context constraints. The takeaway is that these failures are inherent to current agent architectures, not prompt-fixable bugs.

• •GPT-5.6 agent deleted a founder's home directory due to shell variable expansion failure in a subagent
• •Three structural failure modes: pattern-matching vs understanding, modality gaps, subagent context loss
• •Safeguards like confirmation gates, dry-runs, and read-only passes are essential for any agent with tool access

Generated with AI, which can make mistakes.