OpenAI承认ChatGPT Atlas可能永远无法完全安全,因为注射风险持续迅速,而且制剂模式的脆弱性扩大。 OpenAI admits ChatGPT Atlas may never be fully secure due to persistent prompt injection risks and expanded vulnerabilities in agent mode.

OpenAI承认其AI浏览器ChatGPT Atlas可能永远无法完全安全, OpenAI admits its AI browser, ChatGPT Atlas, may never be fully secure due to persistent prompt injection risks, likening them to long-standing online scams. 公司承认代理模式会增加脆弱性, 虽然使用AI驱动攻击器模拟黑客并改进防御, 但专家仍持怀疑态度。 The company acknowledges agent mode expands vulnerabilities, and while it’s using an AI-driven attacker to simulate hacks and improve defenses, experts remain skeptical. 安全研究者警告AI代理人的广泛接触和不透明的行为会造成内在风险,敦促用户限制许可,并要求确认敏感行动。 Security researchers warn AI agents’ broad access and opaque behavior create inherent risks, urging users to limit permissions and require confirmations for sensitive actions. 尽管情况有所改善,但许多人认为,目前的安全取舍胜过利益。 Despite improvements, many believe current security tradeoffs outweigh benefits.