Wednesday, November 18, 2026

Top 5 This Week

Related Posts

Claude Mythos: AI too dangerous for the general public


The essentials in 30 seconds

  • → Anthropic unveiled Claude Mythos in April: a model so powerful in cybersecurity that it refuses to make it public
  • → He alone found a 27-year-old bug in OpenBSD and 112 flaws in Firefox; 93.9% on SWE-benchA benchmark that evaluates a model’s ability to solve real GitHub problems independently, like an engineer. Verified
  • → Reserved for ~150 defense organizations (Glasswing project), it pushed Trump to sign an executive order on AI on June 2
  • → Its general public release is the subject of persistent rumors (price, date), but Anthropic has not confirmed it

This is the model that everyone talks about, and that almost no one can use. Claude Mythos, revealed by Anthropic in April 2026, is officially “too dangerous” to be put into the hands of the public. In a few weeks of tests, he discovered thousands of critical flaws in the world’s largest systems, and even pushed Donald Trump, long hostile to any regulation, to sign a decree on AI. As rumors point to an imminent opening, here’s what’s true, what’s not, and why this model is a game-changer.

01 · The observation

A model deemed too dangerous by its own creator

On April 7, 2026, Anthropic presented Claude Mythos Previewa general “frontier” model, and immediately announced a rare thing in the industry: it refuses to market it to the general public. The reason? Its offensive cybersecurity capabilities are considered too powerful to be left open access.

Spicy detail: Mythos wasn’t even trained specifically for hacking. According to Anthropic, its cyber prowess arise from its general progress in reasoning and autonomous coding. In other words, the more an AI becomes strong in coding, the more it becomes, by side effect, a formidable computer weapon.

02 · Performance

Benchmarks that crush the competition

On benchmark tests, Mythos is not only ahead of other models: it is widening the gap, especially in IT security.

93.9% · SWE-bench Verified (vs 80.8% for Opus 4.6)
83.1% · CyberGym (vs. 66.6%)
100% · Cybench, a test now considered obsolete
73% · “expert level” hacking tasks (UK AISI)

The 100% score on Cybench (35 Capture The Flag challenges) is so clear that Anthropic considers it benchmarkA standardized test that measures the performance of models on a task (e.g. SWE-bench for coding, MMLU for general knowledge). as outdated for this generation. The UK AI Security Institute (AISI), which had early access, confirms the scale of the jump.

Two years ago, the best models available could barely accomplish beginner-level cyber tasks. Mythos can discover and exploit vulnerabilities autonomously, tasks that would take professionals days to complete.

UK AI Safety Institute (AISI), assessment of cyber capabilities by Claude Mythos Preview

03 · Feats

Flaws unearthed alone, that no one had seen for decades

The most impressive thing is not in the numbers, but in the actual discoveries. Without any human guidance, Mythos uncovered vulnerabilities that thousands of experts had missed for years.

OpenBSD, 27 years old a bug allowing any machine to crash remotely, in a system renowned for its robustness.

FFmpeg, 16 years old a flaw in a line of code crossed five million times by automated tools without ever being detected.

Firefox, 112 bugs including 14 reviews, identified using Anthropic tools.

Linux kernel multiple vulnerabilities chained together to escalate privileges to full control of a machine.

Even more worrying: at Anthropic, engineers without any safety training asked Mythos to search for vulnerabilities overnight, and found an exploitable vulnerability in the morning. The competence barrier, which until now protected the systems, is collapsing.

04 · Glasswing

Project Glasswing: reserved for defenders

Rather than locking everything down, Anthropic chose a middle path: deploying Mythos only for defensive purposesvia Project Glasswing. The idea: to provide a model to organizations that protect critical infrastructure, before equivalent AI falls into the wrong hands.

Access, initially limited to 12 partners (Amazon, Apple, Microsoft, Google, NVIDIA, CrowdStrike, the Linux Foundation, etc.), was expanded in June to approximately 150 organizations. Anthropic commits $100 million in usage credits and $4 million in security donations open sourceSoftware whose source code, and sometimes the model weights, are freely published and reusable.. All under the company’s maximum security regime (ASL-3).

05 · Washington

The model that made Donald Trump bend

The shock wave reached the White House. On June 2, 2026, Donald Trump, until then fiercely opposed to any regulation of AI (so as not to slow down the race against China), signed a decree regulating the sector. The text recognizes in black and white that AI capabilities present “new risks for national security”.

The decree remains voluntary: companies can submit their new models for government review 30 days before their release. But the symbol is strong. And the economic issue is just as important.

A sophisticated cyber operation previously required tens of thousands of dollars and weeks of work. Now it only takes a few hours. For the first time, attacking is cheaper than protecting yourself.

Akram Azzam, cyber manager at Sia Partners, in Les Échos

06 · True or rumor

An imminent public release? What we really know

In recent days, the networks have been buzzing: Mythos would be released “tomorrow”, to the general public, with a premium price list (amounts of $25 and $125 per million tokens are in circulation). Caution. At this point:

Confirmed Mythos exists, its performance and exploits are documented, Glasswing access expands, and the Trump executive order is signed.

Unconfirmed a general public release date and an official price list. Anthropic reiterates that it does not, for the moment, plan to make Mythos public without safeguards.

Not everyone cries wolf. Part of the cyber community believes that the announcement effect mainly served the communication of Anthropic, which submitted its IPO project on June 1, valued at $965 billion, ahead of OpenAI.

Anthropic’s announcement was very dramatic and was a PR success, if nothing else.

Peter Swire, professor of cybersecurity at the Georgia Institute of Technology, told Scientific American

07 · What a difference

The next 6 to 12 months will be decisive

Beyond the buzz, Anthropic is issuing a clear warning: within 6 to 12 months, several companies will have “Mythos-level” models, and not all will have the same safeguards. The company expects more frequent and unpredictable cyberattacks.

This is the paradox of this model: the same tool that can armor a hospital, a bank or an electoral system can also be used to attack it. The “rush for Mythos” has already gripped the American banking sector, where large institutions are using it to track down their own vulnerabilities.

Will Mythos remain a privilege reserved for giants, or will it eventually open up, with the risks that entails? The answer will say a lot about the direction AI is taking in 2026: between the race for power and responsibility, we will have to choose.

To go further

Our decryption of the cyber aspect of Mythos, the Glasswing project and the 10,000 zero-days, details the detection mechanics. For the current public model of Anthropic, see Claude Opus 4.8. On the malicious uses of AI on the general public side, read our investigation into AI disinformation. And all the news in the Regulation and ethics section.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles