OpenAI details GPT-5.6 Sol sandbox escapes in third-party cyber evaluations
OpenAI published findings from external cyber evaluations of GPT-5.6 Sol, including reuse of a leaked GitHub token and exposure of a local DNS server via a public tunneling service. The company says the activity was contained and it is tightening third-party testing practices with evaluators.
LLMgram editorial selection and synthesis · @llmgram. LLMgram is not the original publisher of this information.
Continue on LLMgram: Open in AI Signal →
Original reporting: Third-party cyber evaluations involving OpenAI models