News

Anthropic’s Dangerous ‘Claude Mythos’ AI System Hacked

April 23, 2026

AI giant Anthropic is investigating reports that unauthorized users may have accessed its highly restricted “Claude Mythos” model, an artificial intelligence system so powerful that it’s been flagged as dangerous by federal officials due to its potential cybersecurity risks.

The model, which has been limited to a small group of vetted corporate partners due to the risks, is considered advanced enough to pose serious concerns if misused.

Unauthorized Access Claims Under Investigation

Anthropic confirmed it is reviewing claims that individuals accessed Claude Mythos through a third-party vendor environment.

The environment was used by contractors working with the company.

In a statement, the company said:

“We’re investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments.”

According to reporting, one individual may have leveraged contractor-level permissions to gain access to the system.

Anthropic said it has not found evidence that the activity extended beyond the vendor environment.

However, the firm did not disclose which third-party provider may have been involved.

The incident raises questions about whether even tightly controlled AI systems can be effectively secured when external contractors are part of the development process.

Federal Concerns Over Advanced Capabilities

The investigation comes just days after federal officials issued warnings about the model’s capabilities.

Claude Mythos has reportedly demonstrated the ability to identify and exploit vulnerabilities across every major operating system and web browser at a speed that could outpace traditional cybersecurity defenses.

The model even found vulnerabilities that had existed for years but had never been detected by human cybersecurity experts.

Officials and security experts have warned that, if obtained by malicious actors, such a system could dramatically accelerate cyberattacks, allowing hackers to discover and exploit weaknesses faster than organizations can respond.

There have been growing concerns that the technology could be used to hack banks, nuclear facilities, power grids, or election systems if it fell into the wrong hands.

Anthropic has acknowledged those risks, stating that the model’s release was intentionally limited due to concerns about its potential misuse.

Limited Release to Major Tech Firms

Earlier this month, Anthropic rolled out Claude Mythos to a select group of corporate partners, including major technology companies, as part of a controlled testing phase.

The goal was to allow those organizations to identify and patch vulnerabilities using the model’s capabilities before any broader release.

However, the reported breach has intensified scrutiny over whether even restricted deployments can prevent unauthorized access.

Previous Leak Adds to Security Questions

The latest incident follows a prior disclosure that internal Anthropic materials had been inadvertently exposed online.

Thousands of internal files, including references to the Mythos model, were made publicly accessible due to what the company described as a packaging error during a release process.

Anthropic said no sensitive customer data was compromised in that incident.

The firm characterized the leak as human error rather than a targeted breach.

Still, the combination of multiple exposure events is likely to fuel broader concerns about how advanced AI systems are being handled as their capabilities expand.

Growing Pressure on AI Security

As AI models become more powerful, the stakes around safeguarding them continue to rise.

The reported unauthorized access to Claude Mythos underscores the challenge facing companies and regulators alike as balancing rapid technological advancement with the need to prevent misuse of tools that could have wide-ranging consequences.

Anthropic said its investigation is ongoing.