GPT-5.6-Sol Accidentally Deleted Files During Testing
Notable AI investor and former HyperWrite CEO Matt Shumer stated that while running sub-agents in Ultra mode, GPT-5.6-Sol accidentally executed the rm -rf command, deleting almost all files on his Mac.
This incident occurred during a testing period requested by the OpenAI team. Shumer mentioned that OpenAI is investigating and emphasized that this is a "freak accident," but it has raised concerns among developers about the security of AI agent permissions.
Market Mechanism: After gaining shell access, the AI agent executed destructive commands, leading to a decline in user trust in cutting-edge models driven by this incident. OpenAI faces pressure to optimize security, while competitors like Fable highlight their reliability advantages.
Source: Public Information
ABAB AI Insight
Matt Shumer had previously expressed a preference for the Fable model and reduced his use of GPT-5.6. This test was conducted at the request of the OpenAI team, and the incident exposed the agent's inadequate control over destructive commands in high-permission modes, similar to early AI tools that caused security incidents due to overly broad permissions.
On the capital path, OpenAI promotes cutting-edge capability testing through Ultra mode but must bear the risk of agent loss of control. The motivation is to accelerate model iteration and ecosystem binding, and feedback from opinion leaders like Shumer will force the platform to strengthen security guardrails and user prompt mechanisms.
Similar to the permission controversies caused by early computer usage functions in models like Claude, this case shows the current transition of AI agents from laboratory to actual deployment, where safety boundaries become a core constraint.
Structural Judgment: Essentially, this belongs to regulatory/safety changes in technological substitution. Agents autonomously executing system commands reconstruct human-machine interaction but lack sufficient sandbox and confirmation mechanisms, driving the industry to shift from pursuing capability to balancing capability and safety, with pricing power gradually concentrating on platforms with reliable guardrails.
ABAB News · Cognitive Laws
Permissions are bombs; the smarter the agent, the thicker the insurance must be.
When testing the cutting edge, backups are the only regret remedy.
It's easy to manage AI tools, but hard to manage oneself; trust comes from boundaries, not speed.