Anthropic has introduced two new AI models, Claude Mythos 5 and Claude Fable 5, expanding access while aiming to address cybersecurity risks posed by advanced AI technology. Claude Mythos 5 is available exclusively to a limited group of trusted industry partners, while Claude Fable 5 is broadly released with built-in safeguards to limit misuse.
What Happened
On June 9, 2026, Anthropic announced the release of Claude Mythos 5, a powerful AI model with advanced cybersecurity capabilities, granted only to a select set of technology companies and biology researchers. Alongside this, Anthropic launched Claude Fable 5 for the general public, a version designed with protective mechanisms to prevent exploitation in cybersecurity attacks.
Key Facts
- Claude Mythos 5 builds on the prior Mythos Preview model released in April, offering enhanced abilities, especially in vulnerability discovery and software engineering.
- Claude Fable 5 utilizes the same core technology but restricts responses related to cybersecurity, biology, and chemistry by redirecting sensitive queries to an older AI model, Claude Opus 4.8.
- Anthropic collaborates with the US government in rolling out Claude Mythos 5 under a consortium called Project Glasswing, aimed at preparing stakeholders for the cybersecurity challenges posed by advanced AI.
- Protective measures also reroute queries suspected of AI distillation attempts to guard against misuse.
- Both models come at a higher usage cost, with Claude Fable 5 charging double the rates of Anthropic’s existing public models but less than Mythos Preview.
- Over 1,000 hours of security testing yielded no universal jailbreaks for Claude Fable 5, though caution remains.
Why It Matters
The release highlights the tension between making powerful AI tools broadly available and managing the significant cybersecurity risks they pose. Advanced AI models like Claude Mythos can identify and exploit software vulnerabilities, potentially enabling sophisticated cyberattacks if misused, necessitating guarded access and layered safety mechanisms.
Background
Since April, concerns over AI-driven hacking tools have led Anthropic to limit access to Mythos-level models. Anthropic’s Project Glasswing partners received early access to Mythos Preview to prepare defenses. The company has repeatedly noted that competitors are developing similar capabilities, intensifying the urgency around responsible deployment and access control.
Analysis
Anthropic’s dual-release strategy attempts to balance innovation with safety. By offering a full-capability model to vetted partners and a restricted “safe” version publicly, the company manages risk while gathering broader user feedback. The cautious deployment indicates current technical safeguards are imperfect, prioritizing prevention over unrestricted availability.
Who Is Affected
Selected industry and biology research partners gain advanced cybersecurity AI tools early. General public users access Claude Fable 5, which restricts potentially dangerous queries. Developers and organizations concerned with AI’s cyber threat landscape will monitor how these releases influence security and policy. Markets and competitors are also attentive as AI evolves rapidly.
Reactions / Official Statements
Anthropic officials, including Head of Product Management Diane Penn, have acknowledged the ongoing challenge of safely deploying Mythos-level capabilities. Anthropic continuously refines query classifiers to improve precision in filtering sensitive questions but opts for cautious initial rollout to avoid misuse.
What Remains Unclear
This information was not confirmed in the reviewed sources: the effectiveness of Claude Fable 5’s safeguards over time, detailed criteria for partner selection, and how access will expand beyond the initial trusted groups.
What Comes Next
Anthropic plans to broaden access via its trusted access program once robust safeguards stabilize. Meanwhile, industry groups continue to seek cybersecurity solutions before such AI models become widely exploitable. Ongoing improvements in classifier precision and security testing are expected.
Sources
This article is based on reporting and publicly available information from the following source:
Read more Artificial Intelligence stories on Goka World News.