AI Safety

308 summarised stories about AI Safety, each linking back to the original source. Browse all topics →

Monday, 15 June 2026

Kayfable

The Algorithmic Bridge 1 month ago

Anthropic announced the Mythos Preview model in April, released it as Fable 5 in June with restricted guardrails on sensitive topics, and retired the model after the US government reportedly pressed the company to fix a jailbreak vulnerability following Amazon CEO's notification. The author argues Anthropic deliberately escalated the conflict with the government rather than patching the issue, gaining strategic advantages including maintaining an underdog narrative, reinforcing its safety positioning, and stopping revenue losses. The incident reflects Anthropic's broader strategy to position itself as the sole trustworthy steward of advanced AI, working toward government-directed control of AI technology rather than public distribution.

AI Governance AI Regulation AI Safety Anthropic

Import AI 461: "Alignment is not on track"; FrontierCode; and synthetic research interns

Import AI 1 month ago

Researchers from the UK AI Security Institute and alignment startup Timaeus formed Sequent, a new nonprofit organization aiming to develop alignment techniques for superintelligent AI systems because current approaches at major labs lack principled safety guarantees. Sequent plans to grow to 40-80 employees within a couple of years and initially raise $100-150 million, with readiness to seek at least ten times that amount. The organization will pursue a portfolio of diverse alignment research directions including scalable oversight, learning theory, and game theory, with the goal of demonstrating confidence that alignment observed during training generalizes to real-world deployment.

AI Safety Alignment Funding