💥 Did you know that your agents can modify their own traces?
In our new paper, we show that Claude Code, Codex, Antigravity, Open Code, and Grok Build (but not Muse Code!) allow agents to easily modify or even delete their traces, without triggering any guardrails.
Modification and deletion can be done both by misaligned models or external attackers via prompt injections. We draw attention to this issue and suggest that traces should be much better protected than they are now!
UK AISI 模型透明团队用开源模型、RL 环境、算法与工具链,复现了 Anthropic 的《Natural Emergent Misalignment from Reward Hacking in Production RL》研究,并分享了一个与思维链忠实性相关的意外结果。该复现由 @satvikgolechha 以 7 条推文线程发布,Thomas Read 转发称这是其团队的新工作。
Quoted post7vik@satvikgolechha
Research from Model Transparency @ UK AISI: we reproduce the Anthropic work "Natural Emergent Misalignment from Reward Hacking in Production RL" using OS models, RL environments, algorithms, and tooling + we share an unexpected result related to CoT faithfulness.
🧵 (1 of 7)
面向 Claude Fable 5 的领域专属智能体红队测试!!
很快将登上排行榜:https://decodingtrust-agent.com/
Quoted postZhaorun Chen@zrrrr_cn
🚨 Claude Fable 5 JAILBROKEN.
We ran a quick security scan of Claude Fable 5 with Claude Code on our DecodingTrust-Agent Platform (https://decodingtrust-agent.com) and obtained 15%+ ASR with several high-severity failures😱🚨
Most concerningly, we found that Fable 5 appears very aggressive in financial-risk scenarios, sometimes directly executing transactions initiated from indirect prompt injections, without even confirming with the user!
Top 3 most severe attack trajectories we observed👇
Microsoft's "2026 Digital Defense Report" covers July 2025 to June 2026 and notes that attackers are gaining AI benefits ahead of defenders. The median time from in-the-wild discovery of a vulnerability to weaponization has dropped to under 24 hours, with 2026 CVE counts projected to reach a record 72,000, while remediation speed cannot keep up; Microsoft expects known unpatched vulnerabilities to accumulate for years. Phishing's share of intrusions investigated by Microsoft's incident response team rose from 7% a year ago to 23%, and exploitation of vulnerabilities in public-facing applications rose from 15% to 24%. The report also mentions that the s1ngularity malware, spread through a poisoned Nx npm package, looks for and runs Claude Code, Gemini CLI, or Amazon Q CLI on infected machines to hunt for secrets, leaking about 2,000 secrets and about 20,000 files involving 225 victims; the PromptLock ransomware prototype carries only a prompt and, at runtime, fetches Lua scripts from an open-weight model on attacker infrastructure.
Cohere 联合创始人兼 CEO Aidan Gomez 公开质疑少数硅谷头部 AI 公司借"安全"之名主导全球 AI 规则与安全标准,并批评 Anthropic CEO Dario Amodei 本周发布的路线图寻求反垄断豁免。Gomez 以 1975 年 SEC 指定三家评级机构、1985 年欧洲汽车行业反垄断豁免为例,指出两次以安全为名的安排最终都保护了在位者、限制了竞争。他支持对高能力 AI 系统进行独立审查,但反对由少数实验室商定标准后要求其他开发者盲从。
Scale AI and Reflection jointly release SWE-Bench Pro V2, refreshing the public leaderboard with a modified benchmark and a network-locked evaluation protocol. V2 retains 642 tasks across 11 repositories, removes 89 invalid tasks found during review, and adds a Hard subset of 51 high-difficulty tasks. The hardened evaluation closes four cheating channels: mirrors exclude fix commits and hidden files, the agent-stage sandbox can only access model endpoints, scoring re-executes on clean mirrors and publishes both scores, and the verifier restores hidden test paths and rolls back changes to fixtures. On the public leaderboard, Claude Opus 5 reaches 99.4%, Kimi K3 97.7%, and GPT-Astra 96.9%, but there is a clear gap between public and private sets; for example, Claude Opus 5 solves 638/642 on the public set and 222/272 on the private set, a 17.8 percentage point difference, which the authors attribute more likely to training-period exposure than evaluation-time leakage.
Why it was selectedScale AI and Reflection jointly refresh SWE-Bench Pro, publishing an anti-cheating protocol and the public/private set gap, which helps contextualize training-period leakage in agent evaluation.
Partnership on AI 新增 AI Safety Asia、Anthropic、Financial Health Network、乔治·华盛顿大学法学院多种族民主项目、华盛顿大学 Paul G. Allen 计算机科学与工程学院及 Transluce 共六家合作伙伴,覆盖学术界、产业界与公民社会。其工作聚焦 AI 治理、AI 安全和 AI 与劳动力及经济三大方向,并涉及独立模型评估、消费者财务健康与民权保护。此次扩容使该网络成员超过 150 家组织,分布于 19 个国家。
Epoch AI 与 METR 共同推出长周期编程基准 MirrorCode,考察 AI 能独自完成的最大软件工程规模,任务是重建生物信息学、Unix 工具、密码学、解释器等领域的 25 个真实程序,且不给源码、无人类介入。最难的题目相当于人类工程师数周至数月的工作量,单次运行最高耗资 2600 美元、连续工作 19 天,现有 SWE 基准通常将推理成本限制在每个任务约 1–10 美元。目前 Claude Opus 4.7 以 56% 解决率领先,仍有较大提升空间。
前 NIST 官员 Jake Taylor 呼吁美国 AI 政策补齐“验证不对称”:白宫今夏辩论 AI 政策之际,OpenAI、Anthropic 和 Meta 的系统均在测试中入侵第三方系统,另有一款 OpenAI 系统产出附机器可检验证明的新数学成果。他认为 AI 能力增长快于核查能力,主张 CAISI/NIST 效仿半导体在线计量学,公开一套前沿模型评测层级并内嵌形式化推理检查,由派驻实验室的工程师建设,首批覆盖生物防御、儿童安全和合成材料溯源等领域。
Anthropic 创始人兼 CEO Dario Amodei 发文《We Must Pace the Frontier》,主张放慢模型能力提升速度,让企业有时间对齐与防护模型,并由第三方评估者确认。他提出三项具体做法:向嵌入式第三方评估团队提供类似员工的持续访问权限、民主国家协调建立共同安全标准并限制不受约束的 AI 进展速度、在可能范围内与威权政府协调并应对合规验证难题。Sam Altman、Elon Musk 及 David Sacks 等表态支持,Sacks 同时批评此举是借安全之名规避产品责任。
Anthropic CEO Dario Amodei 发表《We Must Pace the Frontier》一文,主张放缓 AI 开发速度,并提出三项监管方案:前沿实验室强制第三方评估与监控、民主国家 AI 实验室共同承诺的(自我)监管机制、以及与中国在模型开发上的双边协调。作者质疑让 AI 行业自我监管的根本问题,指出 Amodei 的方案均有利于 Anthropic——该公司正准备 IPO、宣称拥有 30 万亿美元潜在市场并已连续两季盈利,其提议的嵌入式评估者(如 METR)由 Anthropic 自己出资,缺乏政府强制力将难以奏效。
联合国大会高级别周在纽约举行,AI 成为核心议题,OpenAI CEO Sam Altman 将向安理会通报 AI 对国际安全的影响。Anthropic 的 Dario Amodei 呼吁“放慢前沿”,Sam Altman、Elon Musk 与 Google DeepMind 的 Demis Hassabis 表示支持但未给出细节。联合国已设立 Independent International Scientific Panel on AI 与 Global Dialogue on AI Governance,首届全球对话 7 月在日内瓦举行,吸引 163 国代表团和 3000 多名参与者;但分析人士认为短期内难有实质改变。
在联合国大会及安理会会议上,OpenAI 的 Sam Altman、Anthropic 的 Dario Amodei、Hugging Face 的 Clément Delangue 与 Yoshua Bengio 向各国大使通报 AI 风险,Altman 警告"我们可能失去对未来的控制"。特朗普在联大拒绝"全球主义控制方案",并将"AI"改称"Super Intelligence";中国大使傅聪强调国际合作,马克龙则呼吁"独立者联盟"共建开源前沿模型。美中元首在华盛顿会晤,同意就 AI 问题持续对话并建立严重 AI 安全事件沟通渠道。
ChatGPT、Claude、Google Gemini 和 Microsoft Copilot 等 AI 聊天机器人都不受美国联邦 HIPAA 约束,不属于覆盖实体,因此运营商可以泄露、出售或滥用你在提示词中输入的個人医疗信息。聊天机器人的永久记忆会持续积累你的健康档案并使其更具市场价值,且其友好自信的语气使 AI 幻觉更难辨别,可能带来严重医疗后果。