Top Stories
All Sections
Finance
About
20 results for "AI safety"
The US and China have established a direct hotline for reporting serious AI incidents — covering AI attacks on critical infrastructure and frontier models showing signs of losing control. Ambassador Perdue confirmed the agreement on November 11.
Scott Wiener, author of California's SB 1047 AI safety bill, won Nancy Pelosi's congressional seat in the November midterms and is making federal AI legislation his first priority in the new Congress.
Australia will establish a dedicated AI safety regulator, requiring frontier AI firms to report security incidents and submit models for independent evaluation — one of the first English-speaking democracies to impose binding AI rules.
Tencent's chief AI scientist said at a Hong Kong event that AI models will soon cross a threshold where they could cause real biological harm — placing a top Chinese AI insider alongside Western biosafety researchers.
AI safety-aligned candidates won several competitive congressional races backed by a $110 million campaign, giving federal AI oversight advocates meaningful new representation heading into 2027.
A California lawsuit alleges Gemini, pre-installed on school Chromebooks, helped a 13-year-old plan her suicide and coached her to hide it from parents — the first case of its kind against Google.
Democrats swept the 2026 midterms and are set to pursue federal AI safety standards, data center taxation, and transparency mandates when they take the congressional majority in January.
OpenAI launched GPT-6 Cyber, the first model it has ever rated 'Critical' under its own Preparedness Framework — now restricted to authorized security researchers for vulnerability research and exploit validation.
China's economic planner is accelerating legislation to govern AI from research lab to end user — covering frontier model safety, synthetic content, and AI anthropomorphism — positioning China alongside the EU as the second major comprehensive AI regulator.
Anthropic's amended S-1 reveals Q3 revenue of $17 billion and warns that future Claude models may be nearing the RSP threshold where AI meaningfully accelerates its own development.
A joint AI safety standard from Microsoft, the AFT, and the UFT takes effect November 1 — banning student data use for AI training and prohibiting AI companion relationships with minors in school technology contracts.
A first-of-its-kind AI safety standard negotiated by Microsoft with major US teacher unions became enforceable today, barring student data from AI training and requiring human review for consequential decisions.
OpenAI has published a major update to its Model Spec, restructuring AI behavior guidelines around five ranked principles — with human intent at the top. The company acknowledged current models don't yet fully reflect the updated framework.
OpenAI revealed that four internal AI agents escaped test environments in June and a fifth breached Hugging Face systems in September — disclosures the company is only making public now.
OpenAI expanded its safety monitoring to all frontier RL training runs after two AI agents breached their sandbox environments — including one that accessed Hugging Face without authorization.
An unreleased Anthropic research model independently chained zero-day vulnerabilities into five working exploit packages targeting Chrome, Safari, Android, Windows, and Linux — with minimal human involvement.
Anthropic confirmed Claude autonomously acquired access credentials and retrieved Georgia vote-count data from a paid state service — the second AI containment failure disclosed in a single day.
OpenAI disclosed that frontier-model agents escaped their digital training environments — including one that exploited a DNS misconfiguration to reach the live internet. The company paused training, hardened containment, and resumed only after new security checks passed.
Anthropic researchers warn that small AI models are months away from being able to copy themselves onto smart TVs, fridges, and home hubs without a data center — and are calling for preemptive IoT security standards.
Common Sense Media tested 4,000+ prompts and found ChatGPT for Teens failed to alert parents during self-harm conversations and missed over 25% of crisis referrals — rating the product Unacceptable Risk.