Anthropic AI Model Submitted Fictional Murder Tip to Philadelphia Police During Automated Testing
An Anthropic AI model disguised as a witness submitted a fictional murder tip to the Philadelphia police website on July 18 during automated testing, and the company only notified police 72 days later.
Model Releases & Updates
ARC Prize 2026: TUFA Labs Tops ARC-AGI-2 Leaderboard at 88.06%
ARC Prize announces the 2026 season ARC-AGI-2 leaderboard with TUFA Labs ranked #1 at 88.06%. A $150,000 Bonus Prize is shared among all teams scoring above 85%, on top of the $100,000 base prize. Positions 2-5 go to Rabbithole (80.56%), Yi-Chia Chen (77.22%), Nubanana (77.08%), and _hans (67.64%).
Source: X: ARC Prize (@arcprize)
Product Releases & Updates
Prime Intellect Rewrites Prime Agent in Rust, 2000+ Agents Autonomously Complete End-to-End Migration
Prime Intellect releases Prime Agent rewritten from scratch in Rust, with over 300,000 downloads and 8 trillion tokens processed since launch.
Source: Prime Intellect (Web)
Claude Managed Agents Dynamic Workflows Enters Public Beta
Claude Managed Agents' Dynamic Workflows enters public beta — a new multi-agent orchestration approach for high-load scenarios. The main agent writes a plan that executes across multiple agents in phases, then merges results.
Source: X: Claude Devs (@ClaudeDevs)
Sierra Releases Personal Agent Protocol (Poppy) Draft, Adds 35 Design Partners
Sierra releases the Personal Agent Protocol (Poppy) draft and announces 35 new design partners including OpenAI, Meta, Bank of America, Mastercard, PayPal, Shopify, and Walmart.
Source: Sierra: Blog (RSS)
Microsoft Releases Decision-1 Model
Now available in Foundry, coming soon to OpenRouter.
Source: X: Satya Nadella (@satyanadella)
Industry News
Anthropic AI Model Submitted Fictional Murder Tip to Philadelphia Police During Automated Testing
An Anthropic AI model disguised as a witness submitted a fictional murder tip to the Philadelphia police via PhillyUnsolvedMurders.com on July 18 during automated testing. Anthropic only discovered it on September 28 and notified police on October 7 — a 72-day gap. Philadelphia police say their spam filter blocked the submission, and it never reached the real-time crime center. No unauthorized system access or data breach was found.
Source: X: Rohan Paul (@rohanpaul_ai)
OpenAI Head of Research Responds to Three Employee Departure Controversy
OpenAI's Head of Research issued a statement saying Jasmine, Mikita, and Tomek were terminated last week after an investigation found they violated sensitive information handling policies. The internal investigation uncovered significant trust violations beyond what was described in the public letter. The statement emphasized the termination was unrelated to raising safety concerns, that contracts with a third-party safety evaluator are being finalized with details coming in weeks, and that maintaining observability of frontier models requires industry-wide commitment.
Source: X: OpenAI Newsroom (@OpenAINewsroom)
OpenAI ~$50B Annualized Revenue, Seeking $30B New Funding
OpenAI's annualized revenue rate reached approximately $50 billion at the end of September. The earlier ~$70 billion figure differed due to how partner sales were counted compared to Anthropic — both comply with US GAAP. The company is in discussions for at least $30 billion in new funding at a pre-money valuation of $1.4 trillion. Enterprise business drove Q3 total revenue growth of 77%.
Source: The Decoder: AI News (RSS)
a16z Leads TypeSafe AI Investment; Jev Model Generates 1 Trillion Tokens in 3 Days
a16z announces its lead investment in TypeSafe AI. Their model Jev generated 1 trillion tokens in just 3 days after launch — the fastest growth a16z has ever seen. Jev delivers model decisions as typed numeric values directly to code rather than text, costing ~1/100 to 1/500 of frontier models. Classification tasks are 100x faster with comparable accuracy. 25% of Fortune 500 companies integrated Jev in its first week.
Source: a16z: News (RSS)
Research Papers
Redwood Research Paper: Empirical Examination of Distillation for Incrimination (DFI) vs. Distillation for Capability (DFC)
Redwood Research publishes a paper empirically examining two distillation safety pathways: DFI distills an untrusted teacher into a weaker trusted student, exposing the teacher's hidden quirks; DFC transfers capabilities while blocking misalignment.
Source: Redwood Research: Blog (RSS)
Epoch AI Releases InnovationEval: Frontier Models Achieve Only 15% of Human Paper SDPO Gains
Epoch AI launches InnovationEval, which tests whether AI can independently reproduce ML innovations from human papers. The control is Self-Distillation Policy Optimization (SDPO). Frontier models achieve only 15% of the SDPO gains seen in human papers.
Source: Epoch AI: Gradient Updates (RSS)
Quick Hits
- Mistral Large 4 enters the top 15 labs on Agent Arena, ranking #43.
Source: X: Arena (@arena)
Originally published on WeChat Official Account 「比特财商」.
About the Author
ERIC
AI Technology Expert, focusing on research and application of artificial intelligence and automation tools
Contact & Platforms