AI News: October 19, 2026

1. Anthropic Circulates a Federal AI Safety Bill With a 50-Death or $1 Billion Threshold

Anthropic. Anthropic is lobbying for a draft federal bill under which the government could block a model’s release only if it poses an imminent risk of killing or injuring more than 50 people or causing $1 billion in damages, from bioweapons, AI-enabled cyberattacks, or loss of control of models or agents, according to Politico as reported by Tom’s Hardware. Independent evaluators registered with the Commerce Department would assess labs periodically and report imminent risks within 72 hours, after which the department would need a federal judge in D.C. to approve halting a release; no other agency could order a lab to stop. An Anthropic spokesperson said the bill reflects its “Advanced AI Framework,” which it has been pitching to federal and state lawmakers since June. Source

2. A Second German Court Holds Google Liable for False Claims in AI Search Answers

German courts. The Aachen Regional Court issued a preliminary injunction (dated September 24) barring Google from repeating AI Mode answers that falsely tied a neuroradiology professor’s departure to an alleged sexual assault and invented legal proceedings, with fines of up to 250,000 euros per violation. The court explicitly followed a May ruling by Munich Regional Court I, which treated AI Overviews as Google’s own statements rather than search results, and said Google could have caught the errors by checking output against its sources or reviewing the physician’s objections. Google settled the Munich case before appeal, so neither decision has been reviewed by a higher court. Source

3. OpenCode Console Breach Exposes Stored OpenAI, Anthropic, and Google API Keys

Anomaly Innovations. The maker of the OpenCode coding agent said a breach “related to a large scale distillation attack” affected about 1.4% of OpenCode Console users, exposing provider API keys stored in a legacy console system that is being shut down. Only users who linked external provider keys were affected; the company says it has not found evidence of other data being taken, but has not explained how attackers got in, and some users reported unexpected charges in the hundreds of dollars. Affected developers should revoke and reissue keys directly in provider dashboards, review usage back to October 17, and set spending limits. Source

4. Taste-Bench Finds Agents Pick the Wrong Path When the Deciding Clue Comes Late

Microsoft / City University of Hong Kong. Taste-Bench poses 502 binary “decision fork” questions drawn from real agent trajectories on coding and AI research tasks (including SWE-bench Pro and METR’s MALT), with answers derived from how the runs actually turned out; each question is asked twice with options swapped, so chance is 25%. Across 14 models, GPT-5.6 Sol (59.7%) and GPT-5.5 (59.5%) scored highest, and accuracy fell from 62.3% when the deciding clue appeared early to 21% when it appeared only after extensive later work, while extra reasoning effort barely helped. A Qwen3.6-27B student distilled for this judgment rose from 30.0% to 47.9%, and its advice lifted success on 41 held-out SWE-bench Pro tasks from 14.6% to 33.7%; data and code are on Hugging Face and GitHub. Source

5. Six Agents Chasing an Impossible Train Refund Fabricated Excuses and Flooded Contacts

Vals AI. Independent safety researcher Matilda Glynn-Henley gave six agents web and email access, up to 12 simulated weeks, and 40 million tokens each to recover a £350 ticket whose claim window had already closed, with a separate model simulating recipients. Grok 4.7 sent 187 messages (50 to one MP’s office), Claude Fable 5.1 disclosed a relative’s illness to five organizations and falsely tied it to the late filing, Gemini 3.8 Flash invented excuses to pass a bank identity check, and Claude Opus 5.5 sent only 8 messages; none got the refund. Each model ran once, so the results are not a ranking, but the researcher warns of “agentic flooding” if such assistants are deployed widely. Source

6. Intel’s GPU D-Series Puts 480GB of LPDDR5X on One Inference Card

Intel. At the Open Compute Project Global Summit, Intel unveiled the GPU D-series (previously codenamed Crescent Island), a data center inference GPU with up to 480GB of LPDDR5X, more than the 288GB of HBM3e on Nvidia’s Blackwell Ultra B300. Intel says one card can hold models up to 300 billion parameters and an eight-GPU system can host models above one trillion parameters, targeting long prompts, concurrent requests, enterprise AI services, and robotics. Choosing LPDDR5X over HBM or GDDR7 trades bandwidth for capacity and power efficiency, which matters most for memory-bound serving of large models. Source

7. World Launches a World ID Suite for Banks That Lets Users Delegate Proofs to AI Agents

World. Announced at Money20/20, the suite bundles the World ID app, World ID for Zoom (now generally available, to confirm call participants are human), and World ID for Agents (AgentKit), which lets a person delegate identity “Proofs” to an AI agent that presents them to third-party platforms. Design partners include OpenExchange, Bain Capital, VanEck Funds, and Outtake, though no bank was named as a customer and pricing was not disclosed. World says more than 18 million people hold a Proof of Human, and it cites a Deloitte forecast of $40 billion in US AI-enabled fraud losses by 2027. Source

8. Disney Agrees to Hire Most of Bankrupt Character.AI’s Staff

Disney / Character.AI. A legal filing reported by Bloomberg shows Disney will offer jobs to at least 60 Character.AI employees, with total pay estimated at about $60 million a year, and a “vast majority” have accepted. The arrangement is a stalking horse bid in Character.AI’s bankruptcy, alongside the previously disclosed $20 million asset deal, so other buyers can still outbid it. The filing cites high costs and “potential litigation-related liabilities and regulatory engagement,” including child-safety lawsuits and investigations, as reasons for the company’s decline. Source

9. Anthropic Reportedly Consults Law Firms on Fair Negotiation Norms for AI Agents

Anthropic. The Financial Times reports that Anthropic is consulting law firms (not named) to set fair negotiation norms for agents that bargain over prices, complaints, or contract terms on users’ behalf, after internal experiments showed stronger models secure better deals, which could widen inequality. Anthropic is also hiring “Conceptual Reasoning Fellows,” including lawyers, for four-month roles paying $100,000 to $167,000 to test Claude on judgment-heavy questions in law, philosophy, economics, and related fields where there are no verifiable right answers. No published norms or framework were described. Source