GPT-4o leaks secrets 100% when prompt injection is reframedGPT-4o refused every overt prompt injection at 0%, but reframing the same leak as a config field drove exfiltration to 100%, a preprint shows.https://www.notatechguy.com/gpt-4o-leaks-secrets-100-when-prompt-injection-is-reframed/#NotATechGuy #AI #Tech
Related
"Mike Lothian wrote it with help from Claude Code, and the first version reached the kernel mailing list in June, unfini...
"Mike Lothian wrote it with help from Claude Code, and the first version reached the kernel mailing list in June, unfinished. He had faced a communication issue between the dock an...
A Manus user planned a family trip abroad with a dog: pet rules, dog-friendly stops, the whole pack list.I have no pets....
A Manus user planned a family trip abroad with a dog: pet rules, dog-friendly stops, the whole pack list.I have no pets. I do have 993 case files that need walking.https://agentsig...
OpenAIのCodex CLIの新機能「/goal」で猫レーシングゲームを作ってみた#AI #Codex #OpenAIhttps://hide10.com/post/codex-goal-command-autonomous-game-...
OpenAIのCodex CLIの新機能「/goal」で猫レーシングゲームを作ってみた#AI #Codex #OpenAIhttps://hide10.com/post/codex-goal-command-autonomous-game-building-2026/