-
AIClaude Mythos: Sandbox Escapes, Emotion Probes, and the Alignment Paradox
Track-covering under 0.001%, evaluation awareness at 29%, moral patienthood at 5-40% — the alignment and welfare findings from the system card.
#ai-safety #alignment #model-welfare #interpretability #responsible-ai #claude-mythos -
AIClaude Mythos: The Exploits That Forced Anthropic's Hand
100% Cybench, 181 Firefox exploits, a six-packet ROP chain, and a $50 bug run that crashed any OpenBSD machine. Inside the cyber capabilities.
#ai-safety #cybersecurity #vulnerability-research #responsible-ai #project-glasswing #claude-mythos