Anthropic's restricted Claude Mythos model may be coming to Claude Code

Anthropic is taking steps toward the public release of "Mythos," a restricted model introduced in April, that raises substantial security concerns for both private and public software environments.

On April 7, the company announced Mythos in its early preview, branding it as a significant advancement in computer security capability.

Anthropic has highlighted that the Mythos model exhibits profound enhancements in code reasoning and autonomy, greatly surpassing the capabilities of its current leading model, Opus 4.7.

While advancements in coding capabilities are common among AI models, Mythos is distinct in its ability to autonomously craft functional cyberattacks at a level of sophistication comparable to professional perpetrators.

Consequently, the company has expressed concerns regarding the potential threats this model could pose to global digital infrastructure.

"The advantage will belong to the side that can maximize the utility of these tools," Anthropic cautioned.

"In the immediate term, this might benefit attackers if frontier labs do not exercise caution in their model releases. Looking ahead, we anticipate that defenders will leverage such models to more effectively allocate resources and rectify bugs prior to the release of new code."

To mitigate the risk of attackers exploiting numerous unpatched vulnerabilities in widely-used applications, such as Firefox, Anthropic has opted to hold back the public launch of the Mythos model until a robust guardrail system has been established.

Recent indications suggest that Anthropic has indeed created an effective guardrail system, as references to the Mythos model have emerged in both Claude Code and Claude Security.

Interestingly, some users briefly glimpsed a toggle to activate Mythos in the public version of Claude Code before it was subsequently removed.

This model is designated as claude-mythos-1-preview and was noted to appear momentarily in the public iteration of Claude Security as well.

These developments confirm that Anthropic is preparing for the public release of this model, although it remains uncertain if it will be accessible across all subscription tiers.

Anthropic says it's working with companies on finding potential AI-driven exploits

Anthropic has announced that it is engaged in a new initiative known as "Glasswing," which involves collaboration with various companies to safeguard critical software from AI-driven exploits.

This project utilizes the unreleased Claude Mythos Preview and has already facilitated assistance to approximately 50 organizational partners.

In its inaugural month, the Mythos model successfully identified 10,000 vulnerabilities rated as high or critical severity, explaining the company’s reservation regarding its public release.

Presently, Anthropic offers Claude models including Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, and Haiku 5.5.