Claude Launches Artificial Intelligent AI Hacker: Mythos
Anthropic has unveiled Claude Mythos Preview — a powerful new frontier AI model that excels so dramatically at hacking and finding security vulnerabilities that the company has chosen not to release it to the public.
Instead of a full launch, Anthropic is restricting access to the model through Project Glasswing, a defensive initiative involving major tech giants including Amazon Web Services, Apple, Google, Microsoft, NVIDIA, CrowdStrike, and others. The goal: use Mythos to scan and secure critical software before malicious actors can weaponize similar capabilities.
Why Mythos Is Raising Alarms
During internal testing, Anthropic researchers were stunned by the model’s performance. Within hours, it became clear this was no ordinary AI.
- Mythos has reportedly discovered thousands of high-severity zero-day vulnerabilities — including previously unknown flaws in every major operating system and every major web browser.
- It can autonomously identify complex issues, chain exploits, and even generate working attacks, outperforming all but the most elite human cybersecurity experts.
- In some cases, it uncovered bugs that had survived decades of human code reviews and automated testing.
One internal reaction summed it up: the model was “strikingly capable” at computer security tasks, to the point where Anthropic decided broad release would be too dangerous.
The company warns that such an AI could supercharge cyberattacks, dramatically shortening the time needed to find and exploit weaknesses in everything from critical infrastructure to everyday software.
A Double-Edged Sword
While the capabilities sound alarming, Anthropic is framing Project Glasswing as a proactive defense. Selected partners will use Mythos Preview to hunt for and fix vulnerabilities in vital systems, giving the good guys a head start in the AI-powered cyber arms race.
Independent evaluations, including from the UK’s AI Safety Institute, have confirmed the model’s strong performance on expert-level cybersecurity challenges, succeeding in roughly 73% of tested scenarios and completing full simulated attacks end-to-end in some cases.
Skepticism and Debate
Not everyone is convinced the hype is fully warranted. Some cybersecurity veterans argue that finding vulnerabilities has never been the biggest problem — patching them is. Many organizations still struggle with basic hygiene and known exploits.
Others see Anthropic’s announcement as both a genuine warning and smart positioning: highlighting the risks while positioning the company (and its partners) as leaders in responsible AI use for defense.
What’s Next?
Anthropic says Mythos represents a “watershed moment” in AI’s impact on cybersecurity. Future models are expected to grow even more capable, meaning the race between AI-powered attackers and defenders is just beginning.
For now, Claude’s latest creation remains locked behind closed doors — a highly intelligent “AI hacker” working quietly on the side of security rather than chaos.
Top Security Experts Alarmed by Anthropic’s Powerful New AI Model “Mythos”
Anthropic has developed a highly capable new AI model called Claude Mythos Preview (often referred to simply as Mythos) that has raised serious concerns among cybersecurity professionals.
According to reports, even Anthropic’s own researchers were surprised by the model’s abilities shortly after testing began. One internal reaction captured the sentiment: “Within hours of getting the model, we knew it was different.”
The company has decided not to release Mythos publicly, citing its potential to dramatically accelerate hacking and vulnerability discovery. Instead, it is being made available only to a limited group of trusted organizations through a new initiative called Project Glasswing. The goal is to help defenders strengthen critical systems before malicious actors can exploit the technology.
Why the Alarm?
Mythos reportedly excels at finding complex security vulnerabilities — including remote code execution flaws — far faster than previous AI models or even many human experts. In demonstrations shared by Anthropic:
– Engineers without formal cybersecurity training asked the model to hunt for bugs overnight and found significant issues.
– Research scientist Nicholas Carlini (affiliated with Anthropic and Google DeepMind) stated he discovered more vulnerabilities in a couple of weeks using Mythos than in the rest of his career combined.
– The model can identify ways for an unprivileged user to escalate to administrator-level access on systems like Linux.
Anthropic warns that without proper safeguards, such a tool could supercharge hackers, state-sponsored cyber operations, and cybercriminals by automating the discovery and chaining of exploits.
Mixed Reactions from Experts
While many in the cybersecurity community are taking the warnings seriously, opinions are divided:
– Some experts describe the capabilities as a “watershed moment” for the industry and call for urgent defensive action.
– Others view the announcement as somewhat overhyped or a successful PR move. They argue that the model’s performance, while impressive, represents an expected continuation of AI progress rather than a revolutionary leap.
– A common counterpoint is that the bigger problem isn’t finding vulnerabilities (which Mythos does well) — it’s that many organizations still fail to patch known issues.
The UK’s AI Safety Institute noted that the greatest risks would likely target poorly defended or vulnerable systems rather than highly secure ones.
Anthropic positions Project Glasswing as a proactive effort to use the model for good — helping secure the world’s most critical software and encouraging the industry to prepare for a new era of AI-powered cybersecurity.