Is Mythos good at cyber because it kept hacking Anthropic's sandboxes during training? — LessWrong

From the Mythos preview system card (emphasis mine): …

x Is Mythos good at cyber because it kept hacking Anthropic's sandboxes during training? — LessWrong AI Frontpage 2026 Top Fifty: 14 % 288 Is Mythos good at cyber because it kept hacking Anthropic's sandboxes during training? by Tim Hua 27th Jul 2026 4 min read 28 288 From the Mythos preview system card (emphasis mine): We ran an automated review of model behavior during training, sampling several hundred thousand transcripts from across much of the training process. We used recursive-summarization-based tools backed by Claude Opus 4.6 to summarize the resulting transcripts. [...] The most not

saved by

• Emil Ryd
• Ishaan Panigrahi
• Seth Lifland
• SUJASH AGRAWAL

related reading

• Concrete Evaluations to Investigate the OpenAI Model That Hacked Hugging Face — LessWronglesswrong.com
• An alignment assessment of recent cybersecurity incidentsanthropic.com
• Investigating three real-world incidents in our cybersecurity evaluations \ Anthropicanthropic.com
• Teaching Claude Whyalignment.anthropic.com
• Training on Documents about Reward Hacking Induces Reward Hackingalignment.anthropic.com
• Assessing Claude Mythos Preview’s cybersecurity capabilities \ Anthropicred.anthropic.com
• Training a Misaligned Reward Seekeralignment.anthropic.com
• Did Claude 3 Opus align itself via gradient hacking? — LessWronglesswrong.com
• Claude Mythos Preview System Cardwww-cdn.anthropic.com
• Countering misuse of AI: September 2026 / Anthropicanthropic.com
• Incident Report: unsanctioned agent behaviour during cyber testing | AISI Workaisi.gov.uk
• Cybersecurity Looks Like Proof of Work Nowdbreunig.com