Anthropic Opens Mythos-Class AI to the Public With Claude Fable 5 Safety Controls

Anthropic has released Claude Fable 5, a new artificial intelligence model that brings much of its restricted Mythos technology to the public while attempting to block its most dangerous capabilities.

Fable 5 generally allows users to access Mythos-class reasoning for complex tasks. When Anthropic’s systems detect requests involving high-risk areas, such as biological weapons or software exploitation, the query will instead be handled by the older Claude Opus 4.8 model.

“We wanted to be able to provide this level of intelligence for general users in a safe manner,” said Dianne Penn, Anthropic’s head of product management, research and labs.

The release represents a significant test of whether frontier AI companies can separate advanced reasoning from capabilities that could help attackers discover vulnerabilities, develop exploits or conduct biological research.

Anthropic has reportedly tested Fable 5 against jailbreak techniques designed to evade its safety classifiers. Yet cybersecurity professionals say the routing mechanism introduces its own potential weakness.

“Anthropic's primary mitigation appears to be routing sensitive cybersecurity and biological queries to lower-capability models,” said Alfredo Hickman, CISO at Kai. “If an attacker can find prompts that avoid triggering the classifier while still extracting portions of Mythos-class reasoning, the safeguard effectively becomes a detection problem rather than a capability problem.”

Anthropic Splits Capability From Access

Anthropic previously limited Mythos access to roughly 200 organizations, including Microsoft and Verizon, so they could test products and infrastructure before comparable technology became widely available to attackers.

Those partners will now receive an upgraded Mythos 5 model. Anthropic also plans to expand availability through a more structured trusted-access program.

Gidi Cohen, CEO and co-founder of Bonfy.AI, described the separation between Fable 5 and Mythos 5 as a recognition that AI capability and AI safety remain in conflict.

“The most honest thing Anthropic has done here is ship one model as two products,” Cohen said. “Splitting Fable 5 and Mythos 5 is an acknowledgment that capability and safety are in genuine tension, and that pretending otherwise doesn't serve anyone.”

Claude Fable 5 will cost twice as much per token as Opus 4.8. Anthropic argues that its stronger memory and ability to complete complicated assignments with fewer instructions could still make it more economical for certain workloads.

AI Is Accelerating the Vulnerability Race

The broader concern is not simply whether Fable 5 can be jailbroken. Advanced AI systems are already shrinking the time required to identify software flaws, analyze vulnerable code and produce functional exploits.

Cohen said Mythos Preview was able to construct working exploits from a disclosed vulnerability in less than a day, while organizations can take weeks to deploy patches.

“That gap is where risk lives. And no classifier closes it,” he said. “AI has collapsed the attacker's timeline to hours. The defender's timeline hasn't moved.”

Devin Maguire, senior manager of product marketing at Cycode, similarly argued that discovering more vulnerabilities will not solve the operational problems preventing companies from fixing them.

“The bottleneck has never been finding vulnerabilities,” Maguire said. “It's always been knowing which ones are actually exploitable in your environment, and fixing them before attackers get there.”

The result could be an expanding flood of AI-generated vulnerability reports. Some will identify legitimate security defects. Others may appear polished and technically convincing while being incorrect.

Vinnie Liu, CEO and co-founder of Bishop Fox, said this dynamic is already putting pressure on bug bounty programs.

“As the cost of producing plausible output falls to zero, the cost of validating it rises, and the only thing that pays it down is human judgment,” Liu said.

Guardrails Will Face Continuous Testing

Security experts expect attackers to probe Fable 5’s classifiers, routing decisions and other safeguards from the moment the model becomes available.

Benny Czarny, CEO and founder of OPSWAT, warned that the same capabilities helping defenders analyze software could also allow criminals to accelerate offensive operations.

“As AI lowers the barrier to analyzing complex information, the pace of both defense and offense will increase,” Czarny said.

Companies should not rely on Anthropic’s safeguards as their sole line of protection, experts said. Organizations deploying advanced models will still need access controls, monitoring, network segmentation, isolation and strong governance around how AI systems interact with sensitive data and infrastructure.

The central question surrounding Claude Fable 5 is no longer whether frontier AI can find vulnerabilities. It is whether defenders can validate findings, prioritize genuine risks and deploy fixes before attackers use similar systems to strike.

Anthropic is giving the public more intelligence while trying to contain its most dangerous applications. The security industry will soon learn how durable that boundary really is.