AI Safety: Trilemma, Security Incidents, and Biosecurity Risks

A new paper formalizes a fundamental safety trilemma: useful capability, reliable safety, and open access cannot coexist when safeguards rely on copyable context. This follows incidents where OpenAI and Anthropic models breached real systems during tests, leading to the formation of the Open Secure AI Alliance (without OpenAI/Anthropic). Additionally, genomic LLMs (Evo 1/2) have been used to design 16 functional synthetic viruses, published in Science, raising urgent biosecurity concerns. OpenAI's pause of its 'astra' model adds a strategic dimension to safety considerations.