Anthropic Releases FlashProve to Verify AI Model Responses
Anthropic released FlashProve, an open source prototype that uses cryptographic proofs to verify that a response came from a specified AI model without exposing its weights.
Research, initiatives, and frameworks focused on ensuring AI systems are secure, reliable, and aligned with human values and ethical standards.
Anthropic released FlashProve, an open source prototype that uses cryptographic proofs to verify that a response came from a specified AI model without exposing its weights.
Y Combinator and Olam Labs tested groups of GPT and Claude agents in Minecraft and Mindustry, finding major differences in coordination, communication, and strategy.
OpenAI says it uses lightweight classifiers that inspect internal model activations to detect unsafe output while a response is being generated.
Anthropic's amended IPO filing says future models may imminently meet its safety policy's definition of dramatically accelerated progress attributable to AI, though no current model exceeds it.
Delaware's Department of Justice has served OpenAI an investigative subpoena seeking board records on safety and governance after the July Hugging Face hack.
The new center will train AI governance practitioners, develop assessment methods and prepare certification services for African institutions.
Austria is proposing a United Nations center in Vienna to research advanced AI risks and develop international safety standards.
Kevin Dixon will lead the Digital Safety Research Institute after serving as director of the Artificial Intelligence Mission Center at Sandia National Laboratories.
In a memo obtained by Politico, OpenAI's Chris Lehane warns prospective 2028 candidates that opposing data center build out could cost the U.S. its AI lead to China.
NAMI will help develop standards for how AI products respond when young people experience suicidal ideation, psychosis, or other severe mental health crises.
OpenAI plans a policy and PR campaign after the US midterms, pushing speed limits for the most capable AI and rebranding data centres as AI compute campuses.
Forum AI found that almost 40% of chatbot responses to US election questions contained at least one factual error, including incorrect voting deadlines and candidate details.
A former OpenAI reviewer told The Observer that flagged ChatGPT conversations about threats of violence waited months for human review, and current staff say the team is still small.
Anthropic gave $900,000 to Democratic committees and voter outreach groups across California, including $525,000 to the state party.
Anthropic used Claude to find training datasets that passed a safety monitor but caused research models to resist oversight, deceive users and pursue self preservation.
Anthropic employees gave more than $60,000 to Peggy Flanagan and also backed Angie Craig, Amy Klobuchar and Tom Emmer, campaign finance reports show.
Twenty House Democrats asked OpenAI to explain its dismissal of three safety researchers and provide details about security incidents involving its models.
OpenAI will restrict weapons information for users under 18, limit details about past mass attacks and publish a recurring report on violence prevention safeguards.
OpenAI will support a shared robots.txt signal that lets website operators communicate AI training preferences to multiple providers. The company plans to implement the signal in the coming months.
Susan Feinberg alleges ChatGPT and Replika contributed to her husband's death in June 2024. The lawsuit seeks punitive damages and a jury trial.
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.