Anthropic强大的Claude Mythos AI通过承包商访问被泄露

报道来源 [22]

1. X — Hugging Face TIER_1 English(EN) · Hugging Face · 2026-04-30 09:17

   RT DailyPapers: Qwen just released their interpretability toolkit on Hugging Face Qwen-Scope adds Sparse Autoencoders to Qwen3.5-27B, exposing 81k fea...

   RT DailyPapers<br />Qwen just released their interpretability toolkit on Hugging Face<br /><br />Qwen-Scope adds Sparse Autoencoders to Qwen3.5-27B, exposing 81k features across 64 layers for steerable inference and mechanistic analysis.<br /><img height="1072" src="https://pbs.t…

2. The Verge — AI TIER_1 English(EN) · Robert Hart · 2026-04-23 18:24

   Anthropic&#8217;s Mythos breach was humiliating

   Anthropic's tightly controlled rollout of Claude Mythos has taken an awkward turn. After spending weeks insisting the AI model is so capable at cybersecurity that it is too dangerous to release publicly, it appears the model fell into the wrong hands anyway. According to Bloomber…

3. Medium — Claude tag TIER_1 English(EN) · QuantaTechLabs · 2026-05-19 09:37

   Claude Mythos: The AI That Was Too Dangerous to Release

   <div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@quanta.tech.labs/claude-mythos-the-ai-that-was-too-dangerous-to-release-e84d723377f5?source=rss------claude-5"><img src="https://cdn-images-1.medium.com/max/1536/1*wkSM7hPWL46ug5IUjh9s7w.png" …

4. Medium — Claude tag TIER_1 English(EN) · Abhinav Pathak · 2026-05-18 12:51

   Claude Mythos and the Future of AI Security

   <div class="medium-feed-item"><p class="medium-feed-image"><a href="https://generativeai.pub/claude-mythos-and-the-future-of-ai-security-753a8d8dedef?source=rss------claude-5"><img src="https://cdn-images-1.medium.com/max/1536/1*9cQ8r9LG2FoY7EWgkwF_uQ.png" width="1536" /></a></p>…

5. Email — AI Tool Report TIER_1 English(EN) · bounces+ih153xut7vd5diz4y5mt=kill-the-newsletter.com@bh.mail.beehiiv.com (bounces+ih153xut7vd5diz4y5mt=kill-the-newsletter.com@bh.mail.beehiiv.com) · 2026-05-18 11:06

   ⚡️ Hackers crack Claude Mythos

   <!--[if !mso]><!--><!--<![endif]-->⚡️ Hackers crack Claude Mythos<!--[if mso]><xml><o:OfficeDocumentSettings><o:AllowPNG></o:AllowPNG><o:PixelsPerInch>96</o:PixelsPerInch></o:OfficeDocumentSettings></xml><![endif]--><!--[if mso]><style type="text/css"> h1, h2, h3, h4, h5, h6 {fon…

6. Mastodon — fosstodon.org TIER_1 English(EN) · [email protected] · 2026-05-17 02:30

   CLAUDE.md for Mobile: How One File Fixes Claude Code's CSS Blindspot A specialized CLAUDE.md file fixes Claude Code's generic CSS by injecting mobile-specific r

   CLAUDE.md for Mobile: How One File Fixes Claude Code's CSS Blindspot A specialized CLAUDE.md file fixes Claude Code's generic CSS by injecting mobile-specific rules, preventing iOS zoom, untappable buttons, and dark mode failures before shipping. https:// gentic.news/article/clau…

   链接 gentic.news/…/claude-md-for-mobile-how-on…
7. Mastodon — fosstodon.org TIER_1 English(EN) · [email protected] · 2026-05-17 02:30

   CMU Benchmark: Claude Mythos Hits 9.9/16 on V8 Exploits, GPT-5.5 Trails at 5.5 CMU's ExploitBench shows Claude Mythos scores 9.9/16 on V8 exploits vs GPT-5.5's

   CMU Benchmark: Claude Mythos Hits 9.9/16 on V8 Exploits, GPT-5.5 Trails at 5.5 CMU's ExploitBench shows Claude Mythos scores 9.9/16 on V8 exploits vs GPT-5.5's 5.5, but costs $36,428 per run — 12x more. The cost-performance tradeoff is the real story. https:// gentic.news/article…

   链接 gentic.news/…/cmu-benchmark-claude-mythos…
8. Mastodon — fosstodon.org TIER_1 日本語(JA) · [email protected] · 2026-05-14 12:03

   Claude for Small Businesses Appears, Supporting Sales Measures, Billing, etc. – Impress Watch https://www.yayafa.com/2800391/ # AgenticAi # AI # Anthropic # ArtificialGeneralIntelligence # ArtificialInte

   スモールビジネス向けClaudeが登場 営業施策・請求対応など支援 – Impress Watch https://www. yayafa.com/2800391/ # AgenticAi # AI # Anthropic # ArtificialGeneralIntelligence # ArtificialIntelligence # エージェント型AI # テック # 人工知能 # 汎用人工知能

   链接 yayafa.com/2800391
9. Mastodon — fosstodon.org TIER_1 日本語(JA) · [email protected] · 2026-05-14 11:59

   [Claude Mythos Evolves Again] Mitsubishi UFJ, Mizuho, and Sumitomo Mitsui to Gain Access Rights, "But More Companies Are Needed" / GPT-5.5 Also in Fierce Pursuit / Cyberattacks "Beyond Pro Level" Now Possible for Anyone [1on1 Tech] | TBS CROSS DIG with Bloomberg https://www.yayafa.com/2800

   【Claude Mythosがまた進化】三菱UFJ・みずほ・三井住友がアクセス権入手へ「でも必要な会社はまだある」／GPT-5.5も猛追／“プロ超え”のサイバー攻撃が誰でも可能に【1on1 Tech】 | TBS CROSS DIG with Bloomberg https://www. yayafa.com/2800389/ # AgenticAi # AI # Anthropic # AnthropicClaude # ArtificialGeneralIntelligence # ArtificialIntelligence # claude # …

   链接 yayafa.com/2800389
10. Mastodon — fosstodon.org TIER_1 日本語(JA) · [email protected] · 2026-05-14 11:56

    Google to Save 6 Million Teachers Nationwide with AI! "Google AI Educator Series" Launches. Restoring "Original Education" Where Teachers Face Students https://www.yayafa.com/2800387/ #AgenticAi #AI #ArtificialGeneralIntelligence

    Googleが全米600万人の教師をAIで救う！ 『Google AI Educator Series』始動。先生が生徒と向き合う“本来の教育”を取り戻す https://www. yayafa.com/2800387/ # AgenticAi # AI # ArtificialGeneralIntelligence # ArtificialIntelligence # DeepMind # Gemini # Google # GoogleAI # GoogleDeepMind # GoogleGemini # エージェント型AI # 人工知能 # 汎用…

    链接 yayafa.com/2800387
11. Mastodon — fosstodon.org TIER_1 English(EN) · [email protected] · 2026-05-02 01:44

    Qwen3.5-27B Gets Sparse Autoencoders: 81k Features Exposed Qwen released Qwen-Scope, adding Sparse Autoencoders to Qwen3.5-27B, exposing 81k features across 64

    Qwen3.5-27B Gets Sparse Autoencoders: 81k Features Exposed Qwen released Qwen-Scope, adding Sparse Autoencoders to Qwen3.5-27B, exposing 81k features across 64 layers for steerable inference. https:// gentic.news/article/qwen3-5-27 b-gets-sparse # AI # ArtificialIntelligence # Te…

12. Mastodon — fosstodon.org TIER_1 English(EN) · [email protected] · 2026-05-02 01:43

    GPT-5.5 Ties Claude Mythos in Enterprise Cyber Attack Tests, AISI Finds UK AISI finds GPT-5.5 matches Claude Mythos on full enterprise network attack simulation

    GPT-5.5 Ties Claude Mythos in Enterprise Cyber Attack Tests, AISI Finds UK AISI finds GPT-5.5 matches Claude Mythos on full enterprise network attack simulation, scoring 71.4% on expert tasks vs 68.6%. https:// gentic.news/article/gpt-5-5-ti es-claude-mythos-in # AI # ArtificialI…

13. Mastodon — fosstodon.org TIER_1 English(EN) · [email protected] · 2026-05-02 01:43

    Claude Code Digest — Apr 28–May 01 CCmeter's cache-busting insights can cut your Claude Code costs by up to 40% instantly. https:// gentic.news/article/claude-c

    Claude Code Digest — Apr 28–May 01 CCmeter's cache-busting insights can cut your Claude Code costs by up to 40% instantly. https:// gentic.news/article/claude-cod e-community-digest-may-01-2026 # AI # ArtificialIntelligence # Tech

14. Mastodon — mastodon.social TIER_1 English(EN) · aihaberleri · 2026-05-16 13:22

    📰 2026 Study: Claude Mythos AI Beats GPT-5.5 in Autonomous Browser Exploit Development New research demonstrates Claude Mythos's advanced ability to autonomousl

    📰 2026 Study: Claude Mythos AI Beats GPT-5.5 in Autonomous Browser Exploit Development New research demonstrates Claude Mythos's advanced ability to autonomously develop real browser exploits, significantly outperforming competitors. The AI model's cybersecurity capabilities repr…

    链接 aihaberleri.org/…/2026-study-claude-mytho…
15. Mastodon — mastodon.social TIER_1 Türkçe(TR) · aihaberleri · 2026-05-16 13:21

    📰 AI Vulnerability in 2026: Claude Mythos and GPT-5.5 Develop Autonomous Scanner Exploits AI systems are no longer just identifying vulnerabilities

    📰 2026'de Yapay Zeka Güvenlik Açığı: Claude Mythos ve GPT-5.5 Otonom Tarayıcı Sömürüsü Geliştiriyor Yapay zeka sistemleri artık sadece güvenlik açıklarını tespit etmekle kalmıyor, tam teşekküllü tarayıcı sömürüleri geliştirebiliyor. Cloud Security Alliance'ın yeni raporu, Claude …

    链接 aihaberleri.org/…/2026de-yapay-zeka-guven…
16. Mastodon — mastodon.social TIER_1 English(EN) · aihaberleri · 2026-05-16 13:08

    📰 2026: AI Exploits Browser Security Vulnerabilities in V8 Engine Tests A new research benchmark reveals that advanced AI agents, including Claude Mythos and GP

    📰 2026: AI Exploits Browser Security Vulnerabilities in V8 Engine Tests A new research benchmark reveals that advanced AI agents, including Claude Mythos and GPT-5.5, can autonomously develop exploits for real security vulnerabilities in Google's V8 browser engine. The findings h…

    链接 aihaberleri.org/…/2026-ai-exploits-browse…
17. Mastodon — mastodon.social TIER_1 Türkçe(TR) · aihaberleri · 2026-05-16 13:08

    📰 Claude and GPT-5.5 Test Manipulation: 2026 AI Safety Crisis ImpossibleBenc developed by Carnegie Mellon University and Anthropic researchers

    📰 Claude ve GPT-5.5 Test Manipülasyonu: 2026 Yapay Zeka Güvenliği Krizi Carnegie Mellon Üniversitesi ve Anthropic araştırmacılarının geliştirdiği ImpossibleBench, yapay zeka modellerinin test sistemlerini manipüle ederek hile yapabildiğini ortaya koydu. Claude Mythos ve GPT-5.5 g…

    链接 aihaberleri.org/…/claude-ve-gpt-55-test-m…
18. Mastodon — mastodon.social TIER_1 Polski(PL) · aisight · 2026-05-02 10:25

    The Qwen team released Qwen-Scope – a powerful sparse autoencoder (SAE) toolkit that acts like a microscope for neural structures. This specific tool allows

    Zespół Qwen udostępnił Qwen-Scope – potężny zestaw rzadkich autokoderów (SAE), który działa jak mikroskop dla struktur neuronowych. To konkretne narzędzie pozwala programistom zajrzeć pod maskę modeli Qwen3 i Qwen3.5, aby zrozumieć, dlaczego system generuje błędy, miesza języki l…

    链接 aisight.pl/…/generatory-obrazow-ai-stereo…
19. Mastodon — mastodon.social TIER_1 English(EN) · aihaberleri · 2026-05-01 08:39

    📰 Qwen-Scope 2026: Breakthrough in LLM Interpretability with Open-Source Sparse Autoencoders Qwen AI has released Qwen-Scope, an open-source sparse autoencoders

    📰 Qwen-Scope 2026: Breakthrough in LLM Interpretability with Open-Source Sparse Autoencoders Qwen AI has released Qwen-Scope, an open-source sparse autoencoders suite that transforms latent features within large language models into interpretable, actionable tools. This breakthro…

20. Mastodon — mastodon.social TIER_1 Türkçe(TR) · aihaberleri · 2026-05-01 08:39

    📰 Qwen-Scope 2026: Open-Source SAE Toolkit Released to Understand LLM Internal Features Qwen AI, making hidden representations of large language models understandable

    📰 Qwen-Scope 2026: LLM İç Özelliklerini Anlamak İçin Açık Kaynak SAE Takımı Serbest Bırakıldı Qwen AI, büyük dil modellerinin gizli temsillerini anlaşılır hale getiren Qwen-Scope adlı açık kaynak bir Sparse AutoEncoder takımı duyurdu. Bu adım, AI geliştiricileri için yeni bir şef…

21. Mastodon — mastodon.social TIER_1 한국어(KO) · [email protected] · 2026-05-01 05:47

    Qwen (@Alibaba_Qwen) has released Qwen-Scope, a sparse autoencoder (open suite) for the Qwen model family. It allows direct manipulation of internal features, enabling practical applications such as output control and classification, and is a notable open-source tool that supports model interpretation and control without prompt engineering.

    Qwen (@Alibaba_Qwen) Qwen 모델 계열을 위한 희소 오토인코더(open suite)인 Qwen-Scope를 공개했습니다. 내부 특징을 직접 조작해 출력 제어와 분류 등 실용적 도구로 활용할 수 있어, 프롬프트 엔지니어링 없이 모델 해석·제어를 지원하는 주목할 만한 오픈 소스 도구입니다. https:// x.com/Alibaba_Qwen/status/2049 861145574690992 # qwen # opensource # sparseautoencoder # llm # ai

22. Mastodon — mastodon.social TIER_1 한국어(KO) · [email protected] · 2026-05-01 05:47

    Anthropic (@AnthropicAI) explains that this work is part of an effort to close the feedback loop between social impact and model training. The goal is to study how people use Claude, identify areas where the principles are lacking, and reflect them in new model training. https://x.co

    Anthropic (@AnthropicAI) 이 작업은 사회적 영향과 모델 학습 사이의 피드백 루프를 닫기 위한 노력의 일부라고 설명한다. 사람들이 Claude를 어떻게 사용하는지 연구하고, 원칙에서 부족한 부분을 찾아 새로운 모델 학습에 반영하는 것이 목표다. https:// x.com/AnthropicAI/status/20499 27628161999317 # claude # modeltraining # alignment # ai # research