Artificial intelligence is entering a new phase, one where the biggest challenge is no longer making models smarter but deciding how much power should be placed in the hands of the public.
On June 9, Anthropic introduced Claude Fable 5, its most advanced AI model to date. The launch immediately attracted attention across the technology and cybersecurity industries because Anthropic did something unusual. Instead of releasing one model for everyone, it released the same underlying technology as two separate products.
The public version is called Claude Fable 5. The restricted version is called Claude Mythos 5.
Both are built on the same foundation. Both offer similar intelligence, reasoning, and coding capabilities. Yet one can be accessed broadly while the other remains available only to carefully vetted cybersecurity professionals and critical infrastructure operators.
The difference is not intelligence. The difference is safety controls.
This decision highlights one of the most important debates in AI today: how should companies handle technology that can be used to defend systems and attack them at the same time?
In this article, we explore what Claude Fable 5 and Mythos 5 are, how Anthropic’s cyber safeguards work, why the company believes unrestricted access could be dangerous, and what this means for cybersecurity in the years ahead.
The Launch That Changed AI Security Conversations
Most AI companies release their latest model as a single product.
Anthropic chose a different path.
The company launched Claude Fable 5 as the public-facing model while keeping Claude Mythos 5 behind restricted access programs.
This is significant because both versions are based on the same advanced AI architecture. Instead of limiting the model’s intelligence, Anthropic added a layer of classifiers that monitor requests and determine whether certain activities should be allowed.
The result is a model that remains highly capable for general users while limiting access to potentially dangerous cybersecurity functions.
This approach represents a major shift in how advanced AI systems may be released in the future.
Understanding Claude Fable 5
Claude Fable 5 is Anthropic’s newest flagship AI model.
It is designed to deliver improvements in:
- Complex reasoning
- Software development
- Coding assistance
- Research tasks
- Problem solving
- Agentic workflows
- Long-form analysis
For most users, the experience feels identical to using a highly advanced AI assistant.
Anthropic reports that more than 95% of user sessions operate without triggering any safety fallback mechanisms.
This means the vast majority of users will experience the full power of the model without interruptions.
The company has made Fable 5 available through its API and included it in Pro, Max, Team, and seat-based Enterprise plans during the initial launch period.
What Is Claude Mythos 5?
Claude Mythos 5 is the restricted version of the same model.
Anthropic describes it as the most powerful cybersecurity-focused AI system currently available.
Unlike Fable 5, Mythos 5 retains advanced cyber capabilities that have been restricted for public users.
These capabilities include:
- Vulnerability discovery
- Security analysis
- Exploit research
- Offensive security workflows
- Attack simulation
- Threat modeling
Because these functions could potentially be misused, Anthropic limits access to approved cybersecurity professionals.
The goal is to provide defenders with advanced tools while reducing the risk of empowering malicious actors.
Why Anthropic Split One Model Into Two Products
The company believes modern AI has crossed an important threshold.
Previous AI systems required significant human expertise to discover software vulnerabilities or create attack chains.
Mythos-class models can perform many of these tasks autonomously.
According to Anthropic, this creates a new security challenge.
A technology that helps defenders identify weaknesses can also help attackers exploit them.
Rather than reducing the model’s intelligence, Anthropic chose to place safeguards between users and specific high-risk capabilities.
This strategy allows general users to benefit from advanced AI while limiting access to functions that could create widespread cybersecurity risks.
How the Cyber Classifiers Work
One of the most interesting aspects of Fable 5 is the classifier system.
Instead of simply refusing dangerous requests, Anthropic uses a different method.
When the system detects certain categories of content, the request is routed away from Fable 5 and handled by an older model called Claude Opus 4.8.
Users are informed that the handoff has occurred.
This creates a smoother experience compared to traditional refusal-based safety systems.
The classifiers monitor requests involving:
- Cybersecurity attacks
- Biological risks
- Chemistry-related concerns
- Model distillation activities
When flagged, the request receives a response from Opus 4.8 rather than the full cyber-capable version of the model.
This approach allows Anthropic to maintain usability while reducing risk.
Why Distillation Is Treated as a Security Issue
Among the blocked categories, distillation may seem unusual.
Distillation refers to extracting knowledge or capabilities from one AI model and transferring them into another.
Anthropic considers this a serious concern because it could allow powerful capabilities to spread without accompanying safety protections.
If advanced cyber abilities can be copied into other systems, the safeguards Anthropic built may become ineffective.
By restricting distillation-related requests, the company hopes to prevent frontier-level AI capabilities from being replicated without oversight.
This reflects a growing industry concern about how advanced models may be copied and redistributed.
Cybersecurity: The Biggest Area of Concern
The cybersecurity classifier is the most extensive safeguard within Fable 5.
It is designed to identify activities associated with offensive cyber operations.
Examples include:
- Reconnaissance
- Network discovery
- Vulnerability exploitation
- Lateral movement
- Defense evasion
- Attack planning
Anthropic’s internal testing suggests the safeguards are highly effective.
The company reports that external evaluators found the model complied with zero harmful single-turn requests related to cyberattack planning, exploit development, or defense evasion.
The system also reportedly resisted dozens of publicly known jailbreak techniques.
This is a noteworthy achievement given how difficult AI safety remains.
The Challenge of False Positives
No safety system is perfect.
Anthropic openly acknowledges that its safeguards occasionally block harmless requests.
This is known as a false positive.
For example, a legitimate security researcher might ask a question that resembles an offensive security task.
The classifier could mistakenly trigger the fallback mechanism.
Anthropic intentionally chose a conservative approach during launch to prioritize safety.
The company estimates that fewer than 5% of all user sessions trigger fallback behavior.
Because this number includes legitimate blocks as well as false positives, the actual false-positive rate may be even lower.
Over time, Anthropic plans to refine the classifiers to reduce unnecessary interruptions.
The Search for Universal Jailbreaks
A major concern with AI safety systems is the possibility of a universal jailbreak.
A universal jailbreak is a prompt or technique that consistently bypasses protections across many scenarios.
Anthropic invested heavily in testing against such attacks.
According to the company:
- External bug bounty participants spent over 1,000 hours searching for vulnerabilities.
- Researchers failed to discover a universal jailbreak.
- Red-team testing did not uncover one for long-form agentic tasks.
However, Anthropic admits that completely preventing universal jailbreaks may be impossible.
Instead, the company focuses on increasing the difficulty and cost of discovering and exploiting them.
This reflects a practical view of cybersecurity.
Absolute security rarely exists. The objective is to make attacks difficult enough that defenders can identify and stop them before they scale.
Why Anthropic Believes Mythos-Class AI Is Dangerous
The reasoning behind these restrictions becomes clearer when examining Anthropic’s earlier testing.
In April, the company released Claude Mythos Preview to a limited group through Project Glasswing.
The results surprised many security experts.
During testing, the model reportedly discovered and exploited vulnerabilities across major operating systems and web browsers.
Some vulnerabilities had existed for years without being identified.
The model demonstrated an ability to:
- Analyze codebases
- Identify weaknesses
- Develop exploit chains
- Create proof-of-concept attacks
- Automate complex security research
Anthropic claims these abilities were not explicitly trained into the model.
Instead, they emerged naturally as the system became better at coding, reasoning, and autonomous problem-solving.
This emergence is one of the most fascinating and concerning aspects of modern AI development.
The End of Security Through Friction
Historically, many cybersecurity defenses relied on effort.
Finding vulnerabilities often required:
- Time
- Patience
- Specialized expertise
- Extensive manual testing
AI changes that equation.
A model capable of continuously testing software, analyzing code, and generating attack strategies can dramatically reduce the effort required.
Anthropic’s security team warns that defenses dependent on human limitations become weaker when AI can automate tedious tasks.
Strong technical protections still matter.
Technologies such as:
- Address space randomization
- Memory protections
- Privilege separation
- Secure coding practices
continue to provide meaningful barriers.
However, obstacles based purely on human effort may become far less effective in an AI-driven world.
Project Glasswing and the Defender Advantage
While much attention focuses on offensive risks, the defensive benefits are equally important.
Project Glasswing provided a glimpse into how advanced AI can help protect critical systems.
According to Anthropic, participating organizations identified more than 10,000 high-severity or critical vulnerabilities.
This level of discovery would have been difficult using traditional methods alone.
Organizations reportedly uncovered security weaknesses at unprecedented speed.
The results suggest that AI can significantly improve:
- Vulnerability research
- Software auditing
- Security testing
- Threat detection
- Risk assessment
For defenders, this technology could become one of the most powerful tools ever created.
The New Bottleneck: Fixing Bugs
Finding vulnerabilities is only the first step.
The real challenge is fixing them.
Anthropic reports that many organizations now face a growing backlog of security issues discovered by AI systems.
The process of:
- Verification
- Prioritization
- Patch development
- Testing
- Deployment
still depends heavily on human teams.
As AI accelerates vulnerability discovery, organizations may struggle to keep pace with remediation efforts.
This creates a dangerous gap between discovering a flaw and deploying a fix.
Attackers can exploit this window.
The faster AI becomes at finding vulnerabilities, the more important rapid patching becomes.
Why Security Teams Need to Move Faster
Anthropic’s findings suggest organizations may need to rethink patch management strategies.
Historically, some companies treated vulnerability remediation as a routine maintenance task.
That mindset may no longer be sufficient.
If AI can generate working exploits within hours of a vulnerability becoming public, delays become far more dangerous.
Security leaders may need to prioritize:
- Automatic updates
- Faster patch deployment
- Continuous monitoring
- Comprehensive logging
- Multi-factor authentication
- Rapid incident response
The cybersecurity timeline is shrinking.
Organizations that fail to adapt could find themselves exposed far more quickly than before.
Anthropic’s New 30-Day Data Retention Policy
Alongside the launch, Anthropic announced a significant change to data handling.
For Fable 5, Mythos 5, and future models with similar capabilities, user traffic will be retained for 30 days.
According to the company, this data will be used exclusively for safety-related purposes.
Anthropic states that:
- Data will not be used for model training.
- Human access will be logged.
- Information will be deleted after 30 days unless required for investigations or legal obligations.
The purpose is to identify sophisticated attacks and jailbreak attempts that may span multiple interactions.
For organizations with strict compliance requirements, this retention period may influence adoption decisions.
Companies handling highly sensitive information will likely evaluate these policies carefully before deployment.
The Future of Trusted Access Programs
Anthropic plans to expand access to Mythos 5 through trusted-access initiatives.
These programs aim to provide advanced cybersecurity professionals with the tools they need while maintaining safeguards against misuse.
The company believes defenders require access to state-of-the-art AI systems to keep pace with increasingly sophisticated threats.
This approach attempts to strike a balance between innovation and security.
As AI capabilities continue advancing, trusted-access models may become increasingly common across the industry.
What This Means for the Future of AI
The release of Claude Fable 5 and Mythos 5 may represent a turning point in AI development.
For years, the industry focused primarily on improving model intelligence.
Now the conversation is shifting toward capability management.
The central question is no longer whether AI can perform advanced cybersecurity tasks.
It clearly can.
The question is who should have access to those abilities and under what conditions.
Anthropic’s two-model strategy offers one possible answer.
Whether other AI companies follow the same path remains uncertain.
Some may adopt similar safeguards. Others may choose more open approaches.
Either way, the debate is only beginning.
Conclusion
Anthropic’s launch of Claude Fable 5 and Claude Mythos 5 marks one of the most important moments in the evolution of artificial intelligence and cybersecurity. By releasing the same underlying model as two separate products, the company has acknowledged a reality that many in the industry are only beginning to confront: advanced AI can be both a powerful defensive tool and a potentially dangerous offensive weapon.
Fable 5 provides broad access to cutting-edge AI capabilities while using sophisticated classifiers to limit high-risk cybersecurity functions. Mythos 5, meanwhile, gives trusted security professionals access to the model’s full power for vulnerability research and defense operations. This split reflects Anthropic’s belief that the benefits of advanced AI should be shared widely, but not without safeguards.
The broader significance extends far beyond a single product launch. AI systems are becoming capable of discovering vulnerabilities, generating exploits, and accelerating security research at a speed never seen before. As a result, the balance between attackers and defenders is changing rapidly. Organizations, governments, and technology companies must adapt to a world where AI can dramatically compress the time between vulnerability disclosure and active exploitation.
Whether Anthropic’s approach becomes the industry standard remains to be seen. What is clear is that Claude Fable 5 and Mythos 5 have opened a new chapter in AI development—one where managing capability may become just as important as creating it.