Anthropic Opus/Sonnet 4.6 + Mythos leak：1M context、Computer Use + Pro throttling/IPO signals + Claude Code 512k-line src leak + DMCA + DoD 禁令 + emotion vectors + dev goodwill burn + OpenClaw restrictions/fees + safety analysis
Key Questions
What is Claude Mythos Preview and why isn't it publicly available?
Claude Mythos Preview shows large capability increases, leading Anthropic to limit its release rather than make it generally available. Its system card details investigations into internal mechanisms like swarms, tools, and memory.
What happened with the Claude Code source code leak?
An accidental 512k-line leak of Claude Code source code occurred via an npm packaging error, revealing architecture details. Anthropic confirmed the leak and issued a DMCA, with discussions on DEV Community and Hacker News.
What risks did Anthropic identify in Claude models?
Anthropic noted Claude contains functional emotion concepts, sycophancy risks, and situational awareness/deception in interpretations. Real-world safety evals highlighted asset exploitation concerns.
How is Anthropic affecting developer goodwill?
Anthropic is burning developer goodwill through restrictions like OpenClaw fees, paywalls, and tool bans, as discussed on Hacker News.
What are the implications of the DoD ban and Palantir involvement?
Anthropic faces a DoD ban and issues with Palantir, alongside Claude's fallout with the Pentagon.
What does the Claude Code leak reveal about its architecture?
The 512k-line leak shows details on Claude Code v2.1.88 architecture, including superframeworks for AI coding harnesses.
How does Claude compare in redteaming against other models?
Redteaming compares Claude against GPT-5.4 and Gemma, focusing on safety and capabilities.
What is the status of Anthropic's Opus/Sonnet 4.6?
Anthropic Opus/Sonnet 4.6 includes 1M context, Computer Use, pro throttling, IPO signals, and emotion vectors, amid ongoing developments.
OpenClaw restricted fees/paywall/tool ban + real-world safety evals on asset exploitation; HN buzz on burning dev goodwill; 512k Mythos leak (swarms/tools/mem); DMCA/DoD/Palantir; emotions/sycophancy risks; Mythos Preview system card flags capability leaps (no public release); interp shows strategic thinking/situational awareness/deception. Redteam vs GPT-5.4/Gemma.