SourceBloomberg Technology•Stalecollected in 18m
Anthropic Restricts Mythos Access Following US Security Order
💡Major regulatory shift: Anthropic restricts flagship models due to national security orders.
⚡ 30-Second TL;DR
What Changed
Anthropic restricted access to advanced models including Mythos
Why It Matters
This sets a major precedent for how AI companies handle government intervention and national security mandates. It signals a shift toward stricter oversight of frontier model availability.
What To Do Next
Audit your current infrastructure dependencies on Anthropic's frontier models to ensure compliance with emerging regional access restrictions.
Who should care:Founders & Product Leaders
Key Points
- •Anthropic restricted access to advanced models including Mythos
- •The move follows a direct order from the Trump administration
- •National security concerns cited as the primary driver for the restriction
- •Potential regulatory ripple effects for OpenAI, Google, and Meta
🧠 Deep Insight
Background and context from public sources — not the original article. 21 sources cited.
🔑 Enhanced Key Takeaways
- •The US government's order specifically mandated the suspension of access to Anthropic's Fable 5 and Mythos 5 models for all foreign nationals, including those within the United States and foreign national Anthropic employees.
- •The national security concerns cited by the Trump administration stem from a perceived 'jailbreak' vulnerability in Fable 5, which could bypass safeguards and enable the model to identify software vulnerabilities, potentially for sophisticated cyberattacks.
- •This directive represents an unprecedented escalation in US export controls, shifting focus from restricting AI hardware and tools to directly limiting access to advanced AI models themselves.
- •The restriction intensifies an existing dispute between Anthropic and the Trump administration, which previously blacklisted the company for its refusal to allow the US military to use its AI models for domestic surveillance and autonomous weapons systems.
📊 Competitor Analysis▸ Show
While specific pricing and detailed benchmarks for Mythos 5 are limited due to its restricted access, a general comparison with leading competitors can be made based on publicly available information for Anthropic's Claude models and competitor offerings.
| Feature/Category | Anthropic (Claude Fable 5 / Mythos 5) | OpenAI (GPT-5.5) | Google (Gemini 3.1 Pro) |
|---|---|---|---|
| Core Focus | AI safety, steerability, long-context performance, enterprise use, cybersecurity, biology research | General-purpose LLM, frontier reasoning, multimodal input, broad application, deep Microsoft integration | Multimodal applications, mobile integration, context-driven analysis |
| Context Window | 1 million tokens | Not explicitly stated for 5.5, but GPT-5 delivers 1M token context window | Not explicitly stated for 3.1 Pro, but Gemini 2.5 Pro has 200K context window |
| Max Output Tokens | 128,000 tokens | Not explicitly stated for 5.5 | Not explicitly stated for 3.1 Pro |
| Key Differentiator | Constitutional AI for ethical behavior, 'always-on adaptive thinking', state-of-the-art cybersecurity capabilities (Mythos 5) | Reduced hallucinations (GPT-5.5), strong instruction-following, widely adopted API | Strong multimodal capabilities, efficient for mobile, strong reasoning |
| Pricing (per million tokens) | Input: $10, Output: $50 (for Fable 5/Mythos 5) | GPT-5.5: Input: $5, Output: $30; GPT-5.4: Input: $2.50, Output: $15 | Gemini 3.1 Pro: Input: $2, Output: $12 (doubles over 200K input) |
| Availability | Restricted (Fable 5 disabled for foreign nationals, Mythos 5 for Glasswing partners only) | Generally available | Generally available |
🛠️ Technical Deep Dive
- Model Class: Mythos 5 and Claude Fable 5 belong to Anthropic's 'Mythos-class' of models, which are a tier above their previous 'Opus-class' in capability.
- Context Window: Both models feature a 1-million-token context window, a significant increase (5x) over previous models like Opus 4.8.
- Output Tokens: They support a maximum of 128,000 output tokens, which is four times that of Opus 4.8.
- Adaptive Thinking: The models incorporate 'always-on adaptive thinking,' meaning their 'extended thinking' capabilities cannot be disabled, allowing for more deliberate processing.
- Tokenizer: A new tokenizer (Opus 4.7+) is utilized, which can process up to 35% more tokens for the same text compared to earlier models.
- Cybersecurity Capabilities: Mythos 5 is specifically highlighted as state-of-the-art in cybersecurity and biology research. Earlier versions, like Claude Mythos Preview, demonstrated the ability to autonomously find and exploit zero-day vulnerabilities in major operating systems and web browsers.
- Safeguards: Claude Fable 5 is the same underlying model as Mythos 5 but includes robust safeguards to prevent its misuse in high-risk areas like cybersecurity and biology, routing such queries to less capable models like Opus 4.8. Mythos 5, conversely, has these safeguards lifted and is restricted to specialized partners (Project Glasswing).
- Data Retention: Use of Mythos 5 requires acceptance of a 30-day data retention policy for safety monitoring.
- Architectural Approach: Anthropic employs 'Constitutional AI,' a method where a predefined set of principles guides machine learning models to ensure ethical behavior and self-regulation, reducing reliance solely on human feedback.
🔮 Future ImplicationsAI analysis grounded in cited sources
Governments globally will increase direct regulatory intervention and export controls on advanced AI models, not just hardware.
The unprecedented nature of the US government's order to restrict access to AI models themselves sets a strong precedent for other nations to view frontier AI as a critical national security asset requiring direct control.
AI developers will face heightened pressure to implement government-mandated safety features and participate in pre-release security vetting processes.
The Trump administration's recent executive orders establishing frameworks for voluntary pre-release review, coupled with the swift action against Anthropic, underscore the consequences of failing to adequately address government security concerns.
Tensions between AI companies prioritizing open innovation and governments prioritizing national security will intensify.
Anthropic's public disagreement with the government's rationale for the restriction and its ongoing legal battles with the administration highlight a fundamental clash over the control and permissible uses of advanced AI technology.
⏳ Timeline
2021
Anthropic founded by former OpenAI researchers, emphasizing AI safety and interpretability.
2026-03
A federal judge issues a temporary injunction against a Trump administration ban on government use of Anthropic's AI, following the company's refusal to allow military use for surveillance and autonomous weapons.
2026-04
Anthropic releases Claude Mythos Preview through Project Glasswing, showcasing advanced cybersecurity capabilities.
2026-06-02
The Trump administration issues an Executive Order on Advanced AI Innovation and Security, establishing a framework for voluntary pre-release review of frontier AI models.
2026-06-05
President Trump signs National Security Presidential Memorandum 11 (NSPM-11), directing national security agencies to accelerate AI adoption and allowing for contract termination with companies limiting government use.
2026-06-13
The US government issues an export control directive, leading Anthropic to disable access to its Fable 5 and Mythos 5 models for all foreign nationals due to national security concerns over a potential 'jailbreak'.
📎 Sources (21)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Bloomberg Technology ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.