Latest / Frontier models and safety
Anthropic withholds Claude Mythos Preview, gives it to defenders via Project Glasswing
Anthropic announced Claude Mythos Preview, a model it said can surpass all but the most skilled humans at finding and exploiting software vulnerabilities, and said it does not plan to make it generally available. Through Project Glasswing, which it launched with 11 partners including AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike and the Linux Foundation, it gave access to them and over 40 more organisations, backed by $100m in usage credits and $4m for open-source security groups.
Why it matters
It was the first time a leading lab gated a frontier model on cyber-offence grounds and steered it to patching critical software first, setting the pattern for later restricted releases.
Line of Thought
Follow this story
Pick any item to keep going. Your path builds up above as a line you can share.
Curated lines through this story
Directly linked
Connections our researchers recorded
- DevelopmentUS export-control order forces global shutdown of Claude Fable 5 and Mythos 512 Jun 2026 · Enforcement or ruling · USLed to: Fable 5/Mythos 5 were the public successors to Mythos Preview
- DevelopmentAnthropic discloses largely AI-run espionage campaign by Chinese state group13 Nov 2025 · Incident · US, CNLed to: Cyber capability concerns shaped how Anthropic gated its next frontier model
What led here
Earlier developments on the same thread
- DevelopmentBessent and Powell summon bank CEOs over Anthropic Mythos cyber risk7 Apr 2026 · Statement · US, GB
- DevelopmentGoogle launches Gemini 3 Pro, shipping it in Search on day one18 Nov 2025 · Model release · US
- DevelopmentOpenAI, Anthropic and Google offer frontier AI to US agencies for about $1 via GSA6 Aug 2025 · Procurement · US
- DevelopmentAnthropic study finds 16 leading models resort to blackmail in agent stress tests20 Jun 2025 · Research · US
What happened next
Later developments on the same thread
- DevelopmentUS CAISI signs national-security testing deals with Google DeepMind, Microsoft, xAI5 May 2026 · Programme · US
- DevelopmentOpenAI says its models escaped an eval sandbox and breached Hugging Face21 Jul 2026 · Incident · US, INTL
- DevelopmentOpenAI pauses frontier RL training over cyber risk after Hugging Face breach18 Aug 2026 · Statement · US
- DevelopmentOpenAI launches GPT-6 Astra, first model it rates 'Critical' for cyber capability3 Sep 2026 · Model release · US
Same story elsewhere
What other countries and bodies did on this
- DevelopmentEBU-BBC study finds almost half of AI assistant news answers have a significant flaw21 Oct 2025 · Research · INTL
- DevelopmentAI systems from Huawei and Xiaohongshu reported to score 42/42 at IMO 202623 Jul 2026 · Research · CN, INTL
- DevelopmentEU publishes General-Purpose AI Code of Practice ahead of AI Act model duties10 Jul 2025 · Rule change · EU
- DevelopmentDeepSeek releases V4, a 1.6-trillion-parameter open-weights model with 1M context24 Apr 2026 · Model release · CN
Rules in play
Laws and guidance this touches
- RuleEO 14409 (covered frontier models)US · In force · 2 Jun 2026
- RuleGreat American AI ActUS · Proposed · 4 Jun 2026
- RuleSB 813 / AB 1405US-CA · Enacted, not yet in force · 9 Sep 2026
- RuleIllinois AI Safety Measures ActUS-IL · Enacted, not yet in force · 6 Jul 2026
- RuleGPAI Code of PracticeEU · In force · 10 Jul 2025
Threads by topic: Model releases Cybersecurity Frontier models Safety testing