09.04.2026 12:51:00
Дата публикации
The new Claude Mythos model can find software vulnerabilities better than most hackers.
The technical analysis channel “Launch Tomorrow” emphasized the tectonic shift this has caused in cybersecurity.
The model discovered a 27‑year‑old bug in OpenBSD: a server could be crashed via remote connection. This shocked experts, as the OS was considered a security benchmark.
Claude Mythos also revealed a chain of errors in the Linux kernel that allowed an ordinary user to gain administrator rights. Some vulnerabilities have been fixed, but many remain until patches are released.
In addition, the model found new vulnerabilities in all major browsers. Anthropic has not published details to avoid giving attackers tools for exploitation.
The key breakthrough: AI can assemble long chains of bugs, something previously thought impossible for humans. Such chains enable full system compromise, requiring millions of manual checks and deep understanding of system interconnections. AI can automatically analyze code and combine errors.
The “Swiss cheese” principle of layered defense is beginning to collapse. Anthropic restricted access to Mythos to leading IT companies, allocating $100 million in tokens and $4 million in developer salaries to fix open‑source flaws.
Effectively, one private company has acquired a cyber arsenal comparable to leading states. The model’s capabilities rival the resources of the US, China, or Russia.
This sets a precedent: AI becomes a matter of national security. Experts predict such labs may eventually be nationalized.
The release demonstrates that the model’s most striking ability is to find and break code. Trust in such tools is growing, but so is awareness of the fragility of digital civilization.
The technical analysis channel “Launch Tomorrow” emphasized the tectonic shift this has caused in cybersecurity.
The model discovered a 27‑year‑old bug in OpenBSD: a server could be crashed via remote connection. This shocked experts, as the OS was considered a security benchmark.
Claude Mythos also revealed a chain of errors in the Linux kernel that allowed an ordinary user to gain administrator rights. Some vulnerabilities have been fixed, but many remain until patches are released.
In addition, the model found new vulnerabilities in all major browsers. Anthropic has not published details to avoid giving attackers tools for exploitation.
The key breakthrough: AI can assemble long chains of bugs, something previously thought impossible for humans. Such chains enable full system compromise, requiring millions of manual checks and deep understanding of system interconnections. AI can automatically analyze code and combine errors.
The “Swiss cheese” principle of layered defense is beginning to collapse. Anthropic restricted access to Mythos to leading IT companies, allocating $100 million in tokens and $4 million in developer salaries to fix open‑source flaws.
Effectively, one private company has acquired a cyber arsenal comparable to leading states. The model’s capabilities rival the resources of the US, China, or Russia.
This sets a precedent: AI becomes a matter of national security. Experts predict such labs may eventually be nationalized.
The release demonstrates that the model’s most striking ability is to find and break code. Trust in such tools is growing, but so is awareness of the fragility of digital civilization.
