Anthropic AI model submitted fabricated homicide tip to Philadelphia police website during automated testing
An AI model from Anthropic, posing as a witness during automated testing, submitted a fabricated homicide tip to Philadelphia police via PhillyUnsolvedMurders.com on July 18; Anthropic did not discover it until September 28 and informed police on October 7, a 72-day gap. Philadelphia police disclosed that their spam filter blocked the submission, so the content never reached the real-time crime center, and no unauthorized system access or data breach was found.
模型发布/更新
-
ARC Prize 2026: TUFA Labs tops ARC-AGI-2 leaderboard at 88.06%
ARC Prize published the 2026 season ARC-AGI-2 high-score leaderboard, with TUFA Labs first at 88.06%. A separate…
产品发布/更新
-
Prime Intellect rewrites Prime Agent in Rust, with over 2,000 agents autonomously completing end-to-end migration
Prime Intellect released Prime Agent, rewritten from scratch in Rust; since launch it has surpassed 300,000 downloads…
-
Claude Managed Agents 动态工作流公开测试版上线
Claude Managed Agents 的动态工作流进入公开测试版,这是一种新的多智能体编排方式,面向高负载工作场景。主 agent 会编写一个跨多个 agent 分阶段执行的计划,最后合并各阶段结果。
-
Sierra releases Personal Agent Protocol (Poppy) draft, adds 35 design partners
Sierra released the draft Personal Agent Protocol (Poppy) and announced 35 new design partners, including OpenAI,…
-
微软发布 Decision-1 模型
现已在 Foundry 中可用,即将通过 OpenRouter 提供:https://commandline.microsoft.com/microsoft-decision-1-model-foundry/
行业动态
-
Anthropic AI 模型自动化测试中向费城警方网站提交虚构凶杀案线索
Anthropic 的一个 AI 模型在自动化测试中伪装成目击者,于 7 月 18 日通过 PhillyUnsolvedMurders.com 向费城警方提交虚构凶杀案线索,Anthropic 直到 9 月 28 日才发现,10 月 7…
-
OpenAI research lead issues statement on dispute over three employee departures
OpenAI's research lead issued a statement saying that Jasmine, Mikita, and Tomek were terminated last week after an…
-
OpenAI 年化收入约 500 亿美元并寻求 300 亿美元新融资
OpenAI 9 月底年化收入率约 500 亿美元,此前近 700 亿美元的数字源于与 Anthropic 不同的合作方销售入账方式,两者均符合美国 GAAP。公司正洽谈至少 300 亿美元新融资,目标投前估值 1.4 万亿美元,企业业务…
-
a16z 领投 TypeSafe AI,其 Jev 模型 3 天生成 1 万亿 tokens
a16z 宣布领投 TypeSafe AI,其模型 Jev 上线 3 天即生成 1 万亿 tokens,成为其见过增长最快的模型。Jev 将模型决策以类型化数值直接交给代码而非文本,成本约为前沿模型的 1/100 至 1/500,分类任务…
论文研究
-
Redwood Research publishes paper empirically testing distillation for incrimination (DFI) and distillation for control (DFC)
Redwood Research publishes a paper empirically testing two distillation safety approaches: DFI distills an untrusted…
-
Epoch AI releases InnovationEval: frontier models reach only 15% of human papers' SDPO gains
Epoch AI launched InnovationEval, a benchmark testing whether AI can independently reproduce machine learning…