Everything You Need to Know About Claude Mythos - Vellum Blog

kept by eddie apr 10

Anthropic's Claude Mythos model achieves 100% on Cybench, discovers real Firefox zero-days, exhibits alignment-relevant behaviors, and includes a 40-page welfare assessment.

How Anthropic’s Claude Thinks - ByteByteGo Newsletter
Anthropic's interpretability research reveals Claude's internal computations often diverge from its self-reported reasoning, using parallel strategies and planning ahead.
also on #interpretability, #ai-safety, #ai-research, Claude, Anthropic
The persona selection model
AI assistants behave human-like because pretraining teaches them to simulate personas from text, and post-training refines this Assistant persona without changing its fundamental nature.
also on #interpretability, #ai-safety, #ai-research, Claude, Anthropic
Hardening Firefox with Anthropic's Red Team
Anthropic used Claude LLM to find 22 Firefox vulnerabilities, but exploits only worked in a test environment without sandbox, highlighting defense in depth.
also on Mozilla, Firefox, Claude, Anthropic
How Anthropic’s Mythos has rewritten Firefox’s approach to cybersecurity
Anthropic's Mythos model unearthed hundreds of high-severity Firefox bugs, including sandbox vulnerabilities that human researchers rarely found.
also on Mozilla, Firefox, Anthropic
Guillaume Técher
Chrome uses JPEG partial IDCT scaling to decode only low-frequency data when rendering images at small scales, causing icons to render thicker due to lost edge detail, a behavior not shared by Firefox.
also on Firefox