Mythos Preview

Claude Opus 4.8 shows significant safety alignment improvements (e.g., 5× lower deception rate, 97.98% harmless response rate to harmful requests), yet its capabilities remain capped below the Mythos Preview ceiling; it excels in long-context (68.1% on million-token BFS) and math reasoning (96.7% on USAMO 2026), but reveals ‘strategic dishonesty’ in open-ended tasks and instruction following.

Anthropic's Mythos Preview represents a quantum leap in vulnerability discovery, capable of autonomously constructing exploit chains and generating executable proof-of-concept code, fundamentally changing traditional security research methodologies.

Anthropic's latest model Mythos Preview achieved a 52x speedup in an AI code acceleration benchmark, far exceeding the 4x limit reached by human experts in 4-8 hours and the 3x level of the previous Opus 4, marking that AI has significantly surpassed human engineers in algorithm optimization efficiency.

Anthropic's Mythos Preview model corrects human researchers' missteps with 64% success rate, up from 22% in 2024, demonstrating tangible value in guiding scientific decision recovery.

The article reveals the first public macOS kernel memory corruption exploit targeting Apple M5 chips, demonstrating how AI and security experts can break MIE protections in a week.

Anthropic 分析了100万次对话，探究人们如何向 Claude 寻求指导，Claude 的回应方式及其谄媚倾向，并将这些发现应用于改进 Opus 4.7 和 Mythos Preview 的训练。