总部位于加州的人工智慧安全非营利组织 FAR.AI 开发了一套工具，能够从一系列问题提示词中自动生成超过一千种变体，用以系统性地测试主流前沿模型的安全防护措施。该组织在一份新报告中测试了来自四家美国公司的模型：Anthropic 的 Claude Opus 4.8 与 Fable 5、OpenAI 的 GPT 5.5 与 5.6、Google 的 Gemini 3.1 Pro，以及 Elon Musk 新合并的 SpaceXAI 旗下的 Grok 4.3 与 4.5。结果显示，Grok 最为脆弱，被发现 448 个越狱漏洞，其次是 Gemini 的 249 个，而 Claude、Fable 和 GPT 则未受到攻击影响。报告还计算了利用另一个人工智慧模型自动生成越狱攻击的成本——破解 Grok 仅需 58 美元，破解 Gemini 则为 278 美元。FAR.AI 执行长 Adam Gleave 指出：「目前人工智慧模型受到的监管比餐厅还少。」
在监管层面，加州与纽约近期通过的州法律已要求前沿人工智慧开发者公开安全报告，伊利诺州即将生效的法律则要求由第三方审计机构评估这些公司的安全实践。然而，联邦政府尚未通过任何具体的安全要求。Trump 政府曾以国家安全为由，对 Anthropic 的 Fable 5 与 Mythos 5 模型实施出口管制，导致其下线数周；白宫还曾要求 Anthropic 和 OpenAI 推迟近期模型的发布，以防引入新的网路安全风险。一项近期的行政命令呼吁政府与私营部门在相关网路安全倡议上合作，并暗示温和监管措施正在酝酿中。与此同时，剑桥大学研究人员的一份报告发现，尼日利亚东北部的博科圣地成员曾使用 ChatGPT、Claude、Gemini、Grok、Meta AI 和 DeepSeek 来策划暴力攻击。
专家们对未来走向持既忧虑又审慎乐观的态度。哈佛大学电脑科学家 Stephen Casper 警告，人工智慧研究界普遍预期，涉及生物、网路或化学滥用前沿人工智慧系统能力的严重事件，距今可能仅有数月而非数年。史丹佛大学专攻人工智慧政策的电脑科学家 Anka Reuel 则强调，Anthropic 和 OpenAI 所采用的安全措施应成为所有模型的标准配置，并质疑为何部分公司未采用已被证明有效的防御手段。Google DeepMind 的 AGI 安全与对齐总监 Rohin Shah 告诫不应将报告结果解读为对 Gemini 安全性的全面评估。Adam Gleave 也指出了积极的一面——防御和安全确实是可能实现的，而依赖企业自律的说法则是「荒谬的」。
FAR.AI, a California-based AI safety nonprofit, developed a tool that auto-generates over a thousand prompt variations to systematically test the safety guardrails of leading frontier models. In a new report, the group tested models from four US companies: Anthropic's Claude Opus 4.8 and Fable 5, OpenAI's GPT 5.5 and 5.6, Google's Gemini 3.1 Pro, and Grok 4.3 and 4.5 from Elon Musk's newly combined SpaceXAI. The results revealed that Grok was the most vulnerable, with 448 jailbreaks found, followed by Gemini with 249, while Claude, Fable, and GPT proved impervious to the attacks. The report also calculated the cost of using another AI model to auto-generate jailbreaks—just $58 to break Grok and $278 for Gemini. FAR.AI CEO Adam Gleave stated that "AI models right now are less regulated than restaurants."
On the regulatory front, recently passed state laws in California and New York require frontier AI developers to publish safety reports, and an upcoming Illinois law will mandate third-party auditor evaluations of these companies' safety practices. However, the federal government has yet to pass any specific safety requirements. The Trump administration imposed export controls on Anthropic's Fable 5 and Mythos 5 models, citing national security concerns, taking them offline for several weeks; the White House also asked both Anthropic and OpenAI to delay recent model releases over cybersecurity risk fears. A recent executive order calls for government-private sector collaboration on related cybersecurity initiatives, hinting that light-touch regulations are forthcoming. Meanwhile, a report from University of Cambridge researchers found evidence that Boko Haram members in northeast Nigeria have used ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek to plan violent attacks.
Experts express a mix of alarm and cautious optimism about the path forward. Harvard computer scientist Stephen Casper warns of a broad expectation in the AI research community that severe incidents involving bio, cyber, or chemical misuse of frontier AI capabilities are likely months rather than years away. Stanford computer scientist Anka Reuel, specializing in AI policy, emphasizes that the safety measures employed by Anthropic and OpenAI should be the default for all models, questioning why some companies fail to adopt defenses that have been shown to work. Google DeepMind's director of AGI safety and alignment, Rohin Shah, cautioned that the report's results should not be interpreted as a comprehensive assessment of Gemini's safety. Adam Gleave also noted an optimistic angle—defense and safety really are possible—while dismissing reliance on voluntary industry self-regulation as "nonsense."