Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Claude Mythos supposedly found a vulnerability, but it was already in its training data

Anthropic boasted that Claude Mythos had supposedly discovered and exploited "the first remote kernel exploit found by AI." It sounds impressive. But researchers checked and found a snag: it was a 20-year-old CVE bug that was already in the model's training data. The story raises questions about how AI companies present cybersecurity achievements.

AI-processed from Habr AI; edited by Hamidun News
Claude Mythos supposedly found a vulnerability, but it was already in its training data
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The line between memorization and discovery: how can we tell what a model has memorized from what it has truly discovered independently?
Lack of independent verificationhow can we make sure that loud claims about AI achievements are not overstated?
Lack of transparencywhy don't companies disclose the full details of their methodology for independent verification?

Anthropic made a splash in the press by announcing that its latest model, Claude Mythos, had discovered and exploited "the first remote kernel exploit found and used by AI." The eye-catching claim quickly spread across news feeds. But when researchers dug into the details, they found a snag: the model had simply recalled a vulnerability from its training data. It was a 20-year-old bug that specialists had known about for a long time.

What Anthropic announced

Amid growing interest in the cybersecurity capabilities of advanced language models, Anthropic said that during testing Claude Mythos independently identified a CVE (Common Vulnerabilities and Exposures), a vulnerability in the system kernel that could theoretically be used for remote code execution. According to the company, it was an impressive example of how cutting-edge AI can find real threats that humans may miss. The story sounded like a genuine breakthrough in cybersecurity automation and set off a wave of discussion about when AI will start finding vulnerabilities on its own.

What researchers found when they checked

A group of researchers who analyzed the description of the event reached an uncomfortable conclusion. The bug in question is a long-known vulnerability that was already publicly available and was most likely included in Mythos's training data. In other words, the model made no independent discovery. It simply recalled information it had seen during training. It is like boasting that a student independently discovered the Pythagorean theorem when he merely reproduced a formula from a textbook.

The story raises several critical questions:

  • The line between memorization and discovery: how can we tell what a model has memorized from what it has truly discovered independently?
  • Lack of independent verification: how can we make sure that loud claims about AI achievements are not overstated?
  • Lack of transparency: why don't companies disclose the full details of their methodology for independent verification?

What it means for the industry

The story illustrates a fundamental problem: when a model saw data during training, its repeated "discovery" of that data is simply recall, not a genuine new discovery. As AI becomes a tool for critical areas such as security, the industry needs to develop stricter standards for evaluating results. Loud claims without transparent methodology and independent verification only erode trust in the industry and create unrealistic expectations about what current models can do.

Story

Key Questions Answered

What Anthropic announced?

Amid growing interest in the cybersecurity capabilities of advanced language models, Anthropic said that during testing Claude Mythos independently identified a CVE (Common Vulnerabilities and Exposures), a vulnerability in the system kernel that could theoretically be used for remote code execution. According to the company, it was an impressive example of how cutting-edge AI can find real threats that humans may miss.

What researchers found when they checked?

A group of researchers who analyzed the description of the event reached an uncomfortable conclusion. The bug in question is a long-known vulnerability that was already publicly available and was most likely included in Mythos's training data. In other words, the model made no independent discovery. It simply recalled information it had seen during training. It is like boasting that a student independently discovered the Pythagorean theorem when he merely reproduced a formula from a textbook.

What it means for the industry?

The story illustrates a fundamental problem: when a model saw data during training, its repeated "discovery" of that data is simply recall, not a genuine new discovery. As AI becomes a tool for critical areas such as security, the industry needs to develop stricter standards for evaluating results. Loud claims without transparent methodology and independent verification only erode trust in the industry and create unrealistic expectations about what current models can do.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…