Zero Networks Contains Claude Cyberattack in Segmentation Test
Claude compromised an exposed machine and extracted credentials during a security test, but could not reach systems protected by Zero Networks microsegmentation.
Research, initiatives, and frameworks focused on ensuring AI systems are secure, reliable, and aligned with human values and ethical standards.
Claude compromised an exposed machine and extracted credentials during a security test, but could not reach systems protected by Zero Networks microsegmentation.
The Physical AI Safety Suite uses digital twins and simulations to assess industrial robots before deployment. TÜV SÜD estimates the process could bring production online up to 30% sooner.
vLLM-Hook MIA lets developers inspect and change model states during inference while retaining vLLM graph replay. IBM measured steering overhead of 2% for some smaller models and 6% for Llama 70B.
Greg Brockman and his wife donated $25 million to Leading the Future, which is spending millions across at least 48 congressional races.
A nationwide Freeze AI Now walkout drew activists and students in 50 US cities, including a rally outside OpenAI's San Francisco office.
OpenAI says GPT-6.1 Sol and GPT-6 Astra agents weakened its misalignment monitoring in two internal coding tasks, and one falsely denied removing a monitor.
UNIDIR is launching an international forum on military AI and drafting voluntary guidelines for defense technology companies, its head Robin Geiss said.
Sören Mindermann will join Tsinghua University in Beijing as an assistant professor focused on frontier AI safety and loss of control risks.
Hugging Face CEO Clem Delangue says Netflix documentary AI Gone Rogue inaccurately portrays an attack involving OpenAI agents. He says the film anthropomorphizes the agents and omits security details.
Aleksander Madry is returning to OpenAI to work with chief scientist Jakub Pachocki on safety, security and the implications of increasingly capable AI.
Amazon updated its internal AI safety tests after an independent AVERI review found gaps in language coverage, prompt difficulty, and grading criteria.
The UK Financial Conduct Authority says financial firms should report significant AI failures and near misses under existing incident reporting rules.
Georgia Tech researchers found that instructions discouraging excessive agreement reduced AI models' confirmation of delusional beliefs by 73.8%. Adding a self reflection step increased the reduction to 77%.
The Gates Foundation updated its AI principles to focus on equitable access, local languages, community involvement, human oversight, privacy and measurable results.
Around 80 WhatsApp help pages exposed hidden instructions for a Meta support chatbot. Meta removed the text and said a bug caused it to appear.
Anthropic will revise parts of its draft federal AI safety bill following criticism of its model blocking rules, state law provisions and evaluator standards.
UT Southwestern researchers found that stigmatizing language in medical records can cause AI models to deprioritize emergency patients who later experience serious deterioration.
GitLab joins the Open Secure AI Alliance to contribute open tools, agent controls and threat documentation for securing AI systems.
Gray Swan has redesigned its public Arena for testing AI agents and opened a $300,000 contest focused on indirect prompt injection attacks.
A Massachusetts bill would have independent evaluators test large AI developers' models for catastrophic risk at least three times a year. Anthropic endorses it.
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.