# Claude Mythos: Anthropic's AI that threatens skilled jobs


# Claude Mythos: Anthropic's AI deemed too dangerous, and the plan that threatens skilled jobs

**Claude Mythos** is the code name of the artificial intelligence that Anthropic, one of the leaders of the AI market, currently refuses to release. Officially, the model is kept under lock and key for national security reasons. Unofficially, the figures leaking since March 2026 paint a much darker picture: this is not merely an AI that writes better than you, it is a system designed to cover entire careers of developers, analysts and senior executives.

Through very polished marketing, Mythos is presented as the next great AI revolution. Behind the scenes, Anthropic and the big companies of the sector are redefining work, and humans are not necessarily part of the equation any more.

This article retraces Anthropic's trajectory, dissects Claude Mythos' real performance, analyses Anthropic's labour market report and revisits the loosening of its safety policy, to answer one simple question: should we be afraid of Claude Mythos?

{{< admonition abstract "In brief" >}}
- **Claude Mythos** (internal name Capybara): about **10 trillion parameters**, revealed by a leak in March 2026.
- **77.8%** on SWE-bench Pro, **93.9%** on SWE-bench Verified, **83.1%** on CyberGym.
- A **27-year-old vulnerability** discovered in OpenBSD during real tests.
- **1.73 dollars**: the cost for an AI to solve a reverse engineering task that would take an expert 12 hours.
- **75%** of developers' tasks covered by AI, **minus 14%** hiring of 22 to 25-year-olds in exposed sectors.
- **February 2026**: Anthropic removes the obligation to pause training from its safety policy.
{{< /admonition >}}

{{< youtube fI22SoIcn9Q >}}

## What is Claude Mythos?

Claude Mythos is a next-generation language model developed by Anthropic, positioned above the Claude Opus range. Its existence was revealed in **March 2026** following a data leak linked to a configuration vulnerability in the company's CMS, where the code name **Capybara** appeared.

The technical leaks describe a monstrous architecture of about **10 trillion parameters**, presented as a major technological breakthrough exceeding the capabilities of Claude Opus 4.6. Its performance in complex reasoning and offensive cybersecurity is said to be unprecedented, at the cost of very high energy and compute consumption. That offensive capability is precisely what justifies, according to Anthropic, not releasing it.

## From ELIZA to Claude: how Anthropic was born in OpenAI's shadow

To understand Mythos, you have to understand where we come from.

| Year | Milestone |
|---|---|
| 1966 | ELIZA, a simple pattern-matching program, fools people for the length of a conversation. |
| 1990s | Statistical models take over from rule-based systems. |
| 2013 | Word2Vec launches the neural network revolution applied to language. |
| 2017 | Google's paper "Attention Is All You Need" introduces the Transformer and sweeps away the old technical limits. |
| 2018 to 2020 | OpenAI builds the GPT series, up to the explosion of GPT-3 and its 175 billion parameters. |
| 2021 | Dario and Daniela Amodei leave OpenAI with nine researchers and found Anthropic. |

Their fear, in 2021, is that OpenAI would sacrifice safety for profit after Microsoft's massive investment. Anthropic is born with a mission: **Constitutional AI**. Unlike classic human reinforcement, the model evaluates itself against a "constitution" of ethical principles. That is how Claude, the "honest and harmless" AI, is born. But behind this safety facade, a computing beast was being prepared.

## Claude Mythos benchmarks: why it scares

Independent benchmarks are where Mythos becomes genuinely worrying.

| Benchmark | What it measures | Claude Mythos | Competitors |
|---|---|---|---|
| SWE-bench Pro | Complex, real-world coding | **77.8%** | Claude Opus 4.7: 64.3%, GPT-5.5: 58.6% |
| SWE-bench Verified | Solving 500 programming problems | **93.9%** | Near-perfect score |
| CyberGym | Offensive cybersecurity capabilities | **83.1%** | Previous generation: 66.6% |

This is not just theory. In real tests, Mythos discovered a **27-year-old critical vulnerability in OpenBSD**, a system renowned for the rigour of its code audits.

Even more troubling is the economic argument. A human expert takes about **12 hours** to solve a specific reverse engineering challenge. GPT-5.5, Mythos' direct competitor, solved it in **10 minutes** for a cost of **1.73 dollars**. How can a human compete with an hourly cost of a few cents?

{{< admonition tip "Flawfence: agentic AI for your attack surface" >}}
It is in this context that we founded [Flawfence](https://flawfence.com), our attack surface assessment solution. It automates the most tedious and expensive tasks of a pentest: external asset mapping, shadow IT discovery and agentic vulnerability scanning. Not at Mythos' level yet, but thanks to the agentic approach it finds **30% more vulnerabilities** than classic tools.
{{< /admonition >}}

## The impact on jobs: Anthropic's labour market report

In **March 2026**, Anthropic publishes its report on the impact of AI on the labour market. It is a cold shower, all the more credible for coming from the vendor itself.

### Developers, financial and legal analysts on the front line

Unlike factory robots, Mythos targets the **highly qualified**. For developers, **75% of tasks** are now covered by AI in real conditions. Financial and legal analysts rank in the top 10 of the most exposed professions.

### The end of juniors

The cruellest number concerns beginners. Since the launch of ChatGPT, hiring of **22 to 25-year-olds** in these sectors has dropped by **14%**. Companies are not yet laying off seniors en masse, but they are no longer hiring juniors, because AI does their work for a fraction of the price. The result is a drying up of the human pipeline: no juniors today, no seniors tomorrow.

### Claude Code and Claude Design, the workflow layer

Anthropic does not stop at the model. With **Claude Code** and **Claude Design**, the company creates a "workflow layer": the AI no longer helps you code, it owns the environment in which the code is written. A developer spends on average 6 dollars a day through Claude Code for a 30% productivity boost. Profitable for the employer, fatal for the desk next door.

## Safety: Anthropic's pivot

In his essay "The Adolescence of Technology", **Dario Amodei** himself warns that AI could displace **half of office jobs within one to five years**, and calls for an "entente" strategy between democratic nations to manage this power.

Yet in **February 2026**, Anthropic quietly modified its safety policy, the Responsible Scaling Policy. The obligation to "pause" training when risks exceed control was removed, citing market pressure. Safety has become a luxury the company can no longer afford against the competition. Rather ironic, for a company founded precisely on that principle.

{{< admonition warning "The sad plan" >}}
Faced with such a compression of costs and time, Anthropic's strategy pivoted: safety, once its reason for being, has been relegated behind the need to dominate the market. The "sad plan" is not a conspiracy theory, it is an **accounting optimisation**. By automating 75% of the tasks of programmers and analysts, Anthropic drains the pool of human talent in favour of a proprietary infrastructure where the human is no longer the creator, but the mere supervisor of a machine they no longer understand.
{{< /admonition >}}

## Should we be afraid of Claude Mythos?

Given the numbers and the facts, the answer seems to be an unequivocal "yes". We are no longer facing a simple technological evolution, but a **systemic rupture**.

On one side, an unprecedented technical feat: the Capybara tier, its 10 trillion parameters, its ability to unearth 27-year-old vulnerabilities in ultra-secure systems, its 77.8% score on SWE-bench Pro. Mythos no longer merely suggests code: it solves complex architectures that senior engineers would take days to grasp.

On the other, a brutal economic reality. The number to remember is not a benchmark percentage, it is **1.73 dollars**, the cost for the AI to solve a reverse engineering task that would take a human expert 12 hours. In a world where human expertise is sold at that price, what room will be left for our own added value?

## How to prepare, as a professional or a company

1. **Build skills in supervising AI**: designing, framing and verifying an agent's work becomes the differentiating skill.
2. **Invest in what AI does not do yet**: customer relationships, accountability, decision-making under uncertainty.
3. **Keep training juniors**: an organisation that cuts its talent pipeline deprives itself of its future experts.
4. **Treat AI as an attack surface**: autonomous agents hold access and secrets, they must be audited like any other system. We detail these risks in our article on [OpenClaw and AI assistants going off the rails](/en/when-ai-assistants-go-off-the-rails/).
5. **Automate your own defence**: against AI-equipped attackers, continuous monitoring of your attack surface is no longer optional.

## Frequently asked questions about Claude Mythos

### Why does Anthropic not release Claude Mythos?

Officially for national security reasons, because of its offensive cybersecurity capabilities, illustrated by the discovery of a 27-year-old flaw in OpenBSD. The model is said to be reserved for supervised programmes, such as Project Glasswing dedicated to securing critical software.

### Which jobs are most exposed to Claude Mythos?

According to Anthropic's March 2026 report, developers (75% of tasks covered), then financial and legal analysts, are among the most exposed professions. Junior positions are hit first, with a 14% drop in hiring of 22 to 25-year-olds.

### What is the Responsible Scaling Policy?

It is Anthropic's safety policy, which defines the capability thresholds beyond which additional protection measures are required. The February 2026 update removed the obligation to suspend training when risks exceed control capabilities.

### Is Claude Mythos available to companies?

No, not at the time of publishing this article. The accessible models remain the Claude Opus range and its variants, along with the Claude Code and Claude Design tools.

## Key takeaways

Mythos is not only the most powerful AI ever built. It may be the tool that marks the end of work as we know it. The question is no longer whether AI will replace tasks, but who will remain in control, and at what price.

## Sources

- Anthropic, [Labor Market Impacts Report](https://www.anthropic.com/research/labor-market-impacts) (March 2026)
- Anthropic, [Project Glasswing: Securing critical software for the AI era](https://www.anthropic.com/glasswing)
- Anthropic, [Responsible Scaling Policy Updates](https://www.anthropic.com/responsible-scaling-policy)
- Dario Amodei, [The Adolescence of Technology](https://www.darioamodei.com/essay/the-adolescence-of-technology) (2026)
- AI Security Institute (UK), [Our evaluation of OpenAI's GPT-5.5 cyber capabilities](https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-capabilities)
- MindStudio, [Claude Mythos and GPT-5.5 Pass the 'Last Ones' Cyberattack Benchmark](https://www.mindstudio.ai/blog/claude-mythos-gpt-5-5-last-ones-cyberattack-benchmark-results)
- MindStudio, [GPT-5.5 vs Claude Mythos on Cybersecurity](https://www.mindstudio.ai/blog/gpt-5-5-vs-claude-mythos-cybersecurity-benchmark-comparison)
- Medium, [Anthropic's Mythos: what the leaks reveal and what they don't](https://medium.com/@yugank.aman/anthropics-mythos-what-the-leaks-reveal-and-what-they-don-t-as-of-april-26-e6e97486b9c1)
- Medium, [Claude Mythos 5: The First 10-Trillion-Parameter Model](https://medium.com/ai-analytics-diaries/claude-mythos-5-the-first-10-trillion-parameter-model-scaling-laws-hit-a-new-milestone-fa542be336f8)
- Nexford, [What Anthropic's 2026 AI Labor Market Report Means for Your Career](https://www.nexford.edu/insights/what-anthropics-2026-ai-labor-market-report-means-for-your-career)
- Transparency Coalition, [Anthropic abandons safety policy](https://www.transparencycoalition.ai/news/anthropic-abandons-safety-policy-this-is-why-we-work-to-make-ai-safeguards-the-law)
- AI Business, [Anthropic Downgrades its AI Safety Policy Amid Market Pressures](https://aibusiness.com/generative-ai/anthropic-downgrades-its-ai-safety-policy)
- The Expert Community, [Evolution of Large Language Models (2026)](https://theexpertcommunity.com/artificial-intelligence/evolution-of-large-language-models/)
- Wikipedia, [Anthropic](https://en.wikipedia.org/wiki/Anthropic)