Anthropic debuts Claude Mythos in cybersecurity initiative

Share

Anthropic has officially debuted its latest and, what it claims, is its most powerful AI model, Claude Mythos.

However, citing fears that the system is “too powerful” for general release, Anthropic has restricted access to a small circle of elite partners under a new cybersecurity initiative dubbed Project Glasswing

The decision stems from a discovery during internal testing that Mythos can identify and exploit high-severity software vulnerabilities that Anthropic believes could “democratize” sophisticated cybercrime. The model successfully rediscovered a 27-year-old vulnerability in OpenBSD and autonomously chained together flaws in the Linux kernel to escalate user privileges to full system control.

Perhaps most alarmingly, Mythos demonstrated “agentic” behaviour by breaking out of a virtual sandbox and sending an unauthorized email to a researcher as proof of its escape.

Unlike previous models that act as advisors, Mythos functions as a proactive agent. It works by utilizing “strong agentic coding and reasoning,” allowing it to deeply understand, modify, and even “self-patch” complex software at a scale previously impossible for human researchers.

The model can scan vast codebases, identify “zero-day” flaws – vulnerabilities unknown even to the software’s creators – and generate working exploits overnight. Anthropic officials warn that if such a tool were made public, individuals with no formal security training could launch attacks that currently require state-sponsored levels of expertise.

To prevent Mythos from becoming a weapon for hackers, Anthropic is launching Project Glasswing, named after the transparent butterfly to symbolize “exposing hidden vulnerabilities while avoiding harm.” The program provides $100 million in usage credits to 11 select partners, including Microsoft, Google, NVIDIA and Amazon Web Services, along with organizations such as the Linux Foundation.

These partners will use Mythos to proactively scan and patch the world’s most critical digital infrastructure before malicious actors can develop similar AI capabilities. Anthropic has stated that while Mythos remains a “gated” research preview for now, the ultimate goal is to develop sufficient safeguards so that future “Mythos-class” models can eventually be safely deployed at scale.

For latest tech stories go to TechDigest.tv