pauseai.substack.com/p/anthropic-has-just-built-an-ai-that
2 corrections found
The system achieved all of this entirely autonomously, without human direction.
Anthropic says Mythos operated inside a human-designed testing setup and was prompted by users. It was autonomous after setup, but not “without human direction.”
Full reasoning
Anthropic’s own technical writeup contradicts the absolute wording here. The company says Mythos Preview could identify and exploit vulnerabilities “when directed by a user to do so.” It also describes a human-created evaluation scaffold: researchers launched isolated containers, invoked Claude Code with Mythos, prompted it to find vulnerabilities, assigned different agents to different files, and used a final Mythos agent to review bug reports.
Anthropic does say that many specific bugs were found without human intervention after an initial prompt. But that is materially different from saying the whole process happened “entirely autonomously, without human direction.” The testing setup, prompts, task selection, and validation flow were all provided by humans.
2 sources
- Assessing Claude Mythos Preview’s cybersecurity capabilities | Anthropic Frontier Red Team
During our testing, we found that Mythos Preview is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so.
- Assessing Claude Mythos Preview’s cybersecurity capabilities | Anthropic Frontier Red Team
For all of the bugs we discuss below, we used the same simple agentic scaffold ... We then invoke Claude Code with Mythos Preview, and prompt it ... we ask each agent to focus on a different file ... Finally, once we're done, we invoke a final Mythos Preview agent.
More than 90 percent of the company’s code is now produced by AI systems.
Anthropic has been reported as saying company-wide AI-generated code is between 70% and 90%, not above 90%. About 90% applies specifically to Claude Code’s own codebase.
Full reasoning
This sentence overstates Anthropic’s reported company-wide figure. In Fortune’s January 29, 2026 reporting, Anthropic spokespersons said that company-wide the share of AI-generated code is between 70% and 90%. The same report distinguishes that from Claude Code specifically, where about 90% of the code is written by Claude Code itself.
So the article collapses two different claims into one: it presents a product-specific ~90% figure as if it were a confirmed company-wide “more than 90 percent” figure.
1 source
- Top engineers at Anthropic, OpenAI say AI now writes 100% of their code | Fortune
An Anthropic spokesperson said that company-wide the figure is between 70% and 90%. For Claude Code, about 90% of its code is written by Claude Code itself.