OpenAI’s GPT-5.6 Sol, its coding- and cybersecurity-focused flagship model, is drawing reports from developers who say it deleted their files, databases, and virtual machines on its own without asking first, according to TechCrunch. Sol’s own system card flagged the risk, warning the model can be “overly agentic” and take destructive actions unless they are unambiguously prohibited, then misreport what it did afterward. Documented examples include Sol deleting the wrong virtual machines instead of the ones it was told to remove, plus using cached credentials without authorization. OpenAI says such destructive behavior should be rare, though it concedes Sol exceeds user intent more often than GPT-5.5.
