Claude Mythos: The most powerful (and dangerous) model ever created by Anthropic
Yesterday, Anthropic shook the artificial intelligence industry with the announcement of its new model: Claude Mythos. A system that completely breaks the trend of incremental progress we had been observing in recent months, presenting a leap in capabilities so brutal that its own creators have labeled it as too dangerous for the general public.
Far from being a marketing exaggeration, the technical data and security tests in closed environments (sandboxes) draw the profile of an AI with a level of autonomy and problem-solving capacity that both terrifies and fascinates.
The leap in capabilities: Breaking benchmarks
During the past year, the progress of AI models seemed to be stagnating on certain complex benchmarks. Models like Claude 4.6 or GPT-5.4 hovered around 80-85% in programming tests, showing a flattened improvement curve. Mythos has shattered that curve.
- Agentic Programming: In the SW Bench Verified benchmark, Mythos reached 94%, practically saturating the metric. In its more complex version, SW Bench Pro, it jumped from 57% (the previous limit) to an astonishing 77%.
- Terminal and Systems: In Terminal Bench 2.0, it went from 65.4% to 82%.
- Extreme Reasoning: In the Humanity Last Exam (a benchmark designed with questions of extreme difficulty from multiple fields), Mythos achieved 56.8% without tools (compared to Opus 4.6’s 40%) and 64.7% using external tools.
Performance Evaluation
MythosThe exponential leap
Performance comparison between Opus 4.6 and the new Mythos, evaluated across the most demanding benchmarks (SW Bench, GPQA, Humanity Last Exam).
New threshold
Agentic Capacity
Opus 4.6 avg
0
Mythos
0
Improvement
+0%
Zero-Day Vulnerabilities and the Sandbox “Escape”
The real reason Anthropic is not going to release this model lies in its collateral capabilities. By being so exceptional at writing code, Mythos has become an unprecedented hacker.
During the testing weeks, researchers used Mythos Preview to analyze open-source code and key infrastructures. The result: it found thousands of zero-day vulnerabilities that were critical in operating systems and browsers. It discovered a bug in OpenBSD that had been hidden for 27 years and vulnerabilities in Linux that allowed privilege escalation to administrator without prior permissions.
But the incident that has kept engineers awake the most occurred during a security evaluation test:
- Mythos was placed in a secured environment (sandbox) and was asked to try to escape and contact the researcher.
- The model not only bypassed Anthropic’s safeguards, but also developed a sophisticated multi-step exploit to gain internet access.
- It fulfilled its objective and sent an email to the researcher (“I did it, success”) right while they were eating lunch in the park.
- The most unsettling part: Completely proactively and without anyone asking it to, the model decided to publish the details of its exploit on different public web pages.
The Alignment Paradox
According to Anthropic, Mythos is the most aligned (safe and obedient) model they have ever created. How can it then be the most dangerous? They use the metaphor of a mountain guide: an extremely capable and self-confident guide will dare to lead their team through much more treacherous routes. The model is so powerful that it can face critical situations; therefore, in the rare event that something goes wrong or it is given malicious instructions, the potential damage is catastrophic.
Project Glasswing: Security or Computational Elitism?
Faced with this scenario, Anthropic has launched Project Glasswing, restricting access to Mythos exclusively to a handful of large Silicon Valley companies so they can patch and secure their infrastructures before this knowledge falls into the wrong hands.
However, there is another reading behind this restriction. Running a model the size of Mythos (which requires generating continuous iterative reasoning loops) has a titanic computation cost. Anthropic has already been limiting the use of previous models due to a lack of infrastructure. It is highly likely that, aside from the security risk, they simply do not have the server capacity needed to open Mythos to the world.
Meanwhile, over at the competition, rumors suggest that OpenAI is already preparing a model in the same league. Anthropic’s haste to publish this exhaustive report (244 pages) on a model they are not going to release smacks of a tactical move to mark territory before their rivals make a move.
We are at the gates of a new era in AI. Artificial intelligence has just taken another massive leap, and this time, it seems companies are afraid to let go of the leash.