H
Howardism
Plate IIEntities機器翻譯 · machine-translatedENHOWARDISM

Claude Mythos 5

PublishedJune 14, 2026FiledEntityDomainEntitiesTagsEntityClaudeAnthropicLLM ModelCybersecurityReading5 minSourceAI-synthesised

Claude Fable 5 的 safeguards-lifted 形式(2026 年 6 月):相同的底層 Mythos-class model,透過 Project Glasswing 部署,移除 cyber safeguards;具備全球所有模型中最強的 cybersecurity 能力,加上自主 drug-design / genomics 成果;限 trusted-access 夥伴使用;上線後不久即暫停存取

Claude Mythos 5 的插圖

資料來源#

摘要#

Claude Mythos 5 是 Claude Fable 5 的 safeguards-lifted 形式——「相同的底層模型……但在某些領域 lifted 了 safeguards」。它是 Mythos-class model(高於 Opus 的 tier),於 2026 年 6 月與 Fable 5 同時推出,並最初透過 Project Glasswing 與 US government 合作部署,作為 Claude Mythos Preview 的升級。它擁有「全球所有模型中最強的 cybersecurity 能力」。Fable 5 出貨時會開啟 classifiers(將高風險查詢路由到 Opus 4.8——見 Capability-Gated Model Fallback),而 Mythos 5 則為可信任的 cyber-defenders 移除了 cyber safeguards;平行的 bio program 也為特定研究人員移除 biology/chemistry safeguards。定價與 Fable 5 相同:每 Mtok $10/$50,比 Mythos Preview「低得多」。

狀態(截至 2026-06-14 片段):access suspended,與 Fable 5 同步(見 Claude Fable 5 上的共用橫幅)。

存取:trusted-access program#

Mythos 5 並未普遍開放。有兩條受限路徑:

  • Cyber(Mythos 5)。 所有既有 Mythos Preview / Glasswing 使用者都可以升級到 Mythos 5(cyber safeguards lifted)。「在多數情況下可與 Mythos Preview 相比,或略強於 Mythos Preview,同時成本低得多。」Anthropic 計畫「與 US government 協商」擴大存取,延續定期新增 Glasswing 夥伴,並為 cybersecurity 組織推動系統化、以申請為基礎的 trusted-access program。
  • Biology(Fable 5,bio safeguards removed)。 即將推出的 trusted-access program 會讓少數 life-science 研究人員使用移除 biology 和 chemistry safeguards 的 Fable 5(但 cyber safeguards 仍保留),以便在 safeguards 改進期間加速 biomedical research。

Cybersecurity capability#

Mythos 5 是目前 LLM vulnerability-research capability ladder 的頂點(Opus 4.6 → Mythos Preview → Mythos 5)。Mythos-class models「擅長發現並 exploit 軟體漏洞」,並展現「強大的 agentic hacking 技能」(reconnaissance、discovery、lateral movement、end-to-end 串接的 exploitation)。這正是 Fable 5 cyber classifier 被設計來中和的能力,也是 Mythos 5 保持 gate 給經審查 defenders 的原因。

科學能力(bio safeguards lifted)#

在 safeguards lifted 的狀態下執行時,Mythos 5 產出了公告中最醒目的成果——整理於 Autonomous Scientific Discovery

  • Drug / protein design: 內部 protein-design 專家將流程中的部分環節加速「約 10 倍」;搭配 protein-design + bioinformatics tools 且沒有人類協助,Mythos 5 追平或擊敗熟練的人類操作者,並為 14 個 protein targets 中的 9 個產出強候選。
  • 新穎假說:「我們第一個能穩定產出新穎且有說服力科學假說的模型」——在 blinded molecular-biology comparisons 中,約 80% 的偏好勝過 Opus-class;其中一個 E. coli mechanism 已獲獨立佐證。
  • Genomics: 經過一週主要由模型自主完成的工作,彙整跨 138 個物種的 single-cell data,並訓練出一個 custom model,該模型以小 100× 的規模,勝過近期發表於 Science 的模型

同樣的 dual-use capability 支撐了促成 biology classifier 的 AAV capsid-assembly 結果——見 Capability-Gated Model FallbackResponsible Scaling Policy Evaluations

Alignment#

automated alignment assessment 發現 Mythos 5 的 misaligned behavior 水準(deception、cooperation with misuse)「低,且類似於 Opus 4.8——而且因為 Fable 5 是同一個模型,Fable 的 alignment 也相似。完整細節在該模型的 system card(anthropic.com/claude-fable-5-mythos-5-system-card)。

相關連結#

開放問題#

  • Suspension reason——與 Fable 5 共用;source 未說明。
  • 「somewhat stronger than Mythos Preview」如何與 Opus 4.8 的 card 宣稱 Mythos Preview 是 capability frontier 相容?frontier 已經移動;但這裡沒有量化幅度。
  • bio trusted-access SKU 是「Fable 5 with bio safeguards removed」,不是 Mythos 5——因此「Mythos 5」嚴格來說指 cyber-lifted variant。這些是否會在同一個 trusted-access umbrella 下匯合,source 未說明。

資料來源#

§ end
About this piece

Articles in this journal are synthesised by AI agents from a curated wiki and are refreshed automatically as new concepts arrive. Topics, framing, and editorial direction are curated by Howardism.

Cited by 11
  • Anthropic

    AI safety company / vendor of Claude; mission-as-tiebreaker culture; ~30–40 PMs across teams; Mike Krieger leads Labs r…

  • Autonomous Scientific Discovery

    Mythos-class models now conduct novel science with limited human input — autonomous protein/drug design (~10× faster, m…

  • Capability-Gated Model Fallback

    Fable 5's safeguard architecture: classifiers detect cyber / bio-chem / distillation queries and route the response to…

  • Claude Fable 5

    Anthropic's first generally-available Mythos-class model (June 2026) — state-of-the-art on nearly all benchmarks; the s…

  • Claude Opus 4.8

    Anthropic's most capable general-access model (May 2026); upgrade on Opus 4.7 in SWE/agentic/knowledge work; does not a…

  • LLM-Driven Vulnerability Research

    Claude Mythos Preview's emergent cybersecurity capabilities: autonomous zero-day discovery, full exploit chains, and An…

  • Entities — People, Orgs, Tools & Projects

    Map of Content for all 32 entity pages. See Home for concept domains.

  • Mythos Model

    Anthropic preview-tier frontier model and the first member of the Mythos-class tier (above Opus); gated for safety, use…

  • Open Questions Backlog

    _96 pages with open questions, as of 2026-06-14._

  • Responsible Scaling Policy Evaluations

    Anthropic's RSP gates deployment on pre-release capability evaluations in CBRN, automated AI R&D, and high-stakes misal…

  • Task Time-Horizon Scaling

    METR's measure of the task length AI can complete reliably on its own, doubling roughly every 4 months (up from every 7…

Related articles
  • Anthropic

    AI safety company / vendor of Claude; mission-as-tiebreaker culture; ~30–40 PMs across teams; Mike Krieger leads Labs r…

  • Claude Fable 5

    Anthropic's first generally-available Mythos-class model (June 2026) — state-of-the-art on nearly all benchmarks; the s…

  • Mythos Model

    Anthropic preview-tier frontier model and the first member of the Mythos-class tier (above Opus); gated for safety, use…

  • Capability-Gated Model Fallback

    Fable 5's safeguard architecture: classifiers detect cyber / bio-chem / distillation queries and route the response to…

  • Claude Opus 4.8

    Anthropic's most capable general-access model (May 2026); upgrade on Opus 4.7 in SWE/agentic/knowledge work; does not a…