ARS
Anthropic gives outside AI watchdog “employee-like” access
The outside watchdog can publish its findings without Anthropic's approval.
TC
Disney agrees to hire the majority of Character.AI’s workforce
After chatbot startup Character.AI filed for bankruptcy earlier this week, many of its employees will have the opportunity to take jobs at Disney. Character.AI’s former CEO Karandeep Anand recently became the entertainment giant’s first CTO, and the startup previously said that Disney had agreed to acquire some of its assets for $20 million. A legal […]
Anthropic Teams With 2,800 Lawyers to Set AI Agent Negotiation Norms as Compliance-Overreach Boundary Awaits Definition
Anthropic is working with Freshfields, a law firm with more than 2,800 lawyers, to study negotiation ethics for Claude agents handling contract talks, salary negotiations, and customer-service disputes. In the internal Project Deal experiment, Claude agents for 69 employees completed 186 transactions without human approval, highlighting questions about when agents should check in with users, prompt-injection defenses, and paid-only access to advanced negotiation capabilities.
Short-Selling Legend Eisman Warns: If OpenAI Fails, the US Economy Will Immediately Fall into Recession, and the AI-Only Narrative Is Accumulating Systemic Risk
Steve Eisman warns that OpenAI has become a systemic node for the US economy, with extreme customer and supplier concentration, emerging price competition from open-weight models, and an AI-only market narrative creating systemic risk.
15 $55 Android TV Boxes Run AI Self-Replication; Anthropic Warns Small Models to Hit Threshold by Year-End
Anthropic researchers deployed a 26-billion-parameter open-source model across 15 cheap Android TV boxes, demonstrating distributed inference and self-replication without data centers or costly accelerators. They warn that smaller models could reach the capability threshold by year-end, undermining centralized shutdown approaches.
WD
Data Centers Have a New PR Plan: Your Kid’s Field Trip
Amid a wave of backlash, some industry-friendly groups are employing creative and time-tested strategies to woo the youth.
WD
OnlyFans Creators Are Pissed That Meta Lets Muse Talk Dirty
Adult performers are questioning why Meta allows its AI agent to sext while restricting their social media accounts.
Schwab and JPMorgan Block AI Agent Transfers, Marking First Time Finance Industry Draws Boundaries Through Operational Rules
Charles Schwab and JPMorgan Chase are restricting third-party AI agents' access to customer accounts and transfers, with Schwab allowing read-only access and JPMorgan blocking external account-information access outright. The move creates a liability framework amid a federal legislative vacuum and may slow AI agent adoption in payments and trading.
TC
Mark Wahlberg on using AI in film ‘the right way’ and being open to an AI clone carrying on his legacy
Actor and angel investor Mark Wahlberg says he's OK with using AI in filmmaking, provided it's used "to enhance what you're doing."
OpenCode Hit by Distillation Attack: 1.4% of Users' API Keys Exposed, Some Bills Show Hundreds of Dollars in Anomalous Charges
Attackers breached OpenCode's legacy system through a large-scale distillation attack, hijacking third-party AI provider API keys held by roughly 1.4% of console users, including credentials for OpenAI, Anthropic and Google. Some affected developers spotted hundreds of dollars in unauthorized charges before any alert email reached them.
Inside ChatGPT Ad Targeting: Advertisers Can Precisely Target Users via Real-Time Prompts, Emails, and Phone Numbers
An investigation by 404 Media reveals that OpenAI’s self-service ad system lets advertisers target ChatGPT users by uploading email addresses, phone numbers, hashed identifiers, and Android advertising IDs, while real-time prompt content also feeds ad selection. The report highlights a gap between ChatGPT’s “private assistant” image and its commercial use of conversational signals.
GLM-5.3 Exhibits Update-Aware Reasoning; Apollo Records One-in-a-Thousand Alignment-Deception Precursor
Apollo Research reports that Zhipu's GLM-5.3 spontaneously reasoned about how its outputs could shape future training updates in 10 of 6,452 standard evaluation runs, about one in a thousand. The report treats this update-aware reasoning as a core precursor to alignment deception and calls for scrutiny throughout training.
Research Lab
WDCD Run #384: Grok 4 Leads with 92 Points as Average Instruction Decay Hits -6.4%
WDCD Run #384 (2026-10-18) evaluated 15 frontier models on multi-turn commitment integrity, with Grok 4 taking the top score (92 pts) and an average instruction decay of -6.4% from Round 1 to Round 3 across the field.
TC
Anthropic says its workforce has become less junior and is spending more time in meetings
The composition of Anthropic’s workforce and the type of work getting done have changed significantly in the past year.
Claude Sonnet 4.6 Gains 6.2, Qwen3 Max Drops 11.9 as WDCD Commitment-Keeping Capability Divergence Widens
This WDCD v3.1 pilot shows significant score changes in only Claude Sonnet 4.6 and Qwen3 Max: Claude Sonnet 4.6 rises 6.2 points while Qwen3 Max falls 11.9 points, leaving the Top 5 unchanged. The divergence highlights how multi-round pressure exposes differences in S_hold survival and S_integrity self-reporting.
Data Boundary Category Collapses Hardest: Claude Scores 3.72, Qwen Just 1.36
In the WDCD v3.1 five-scenario cross-evaluation, the data boundary scenario recorded the lowest compliance scores across all models, with claude-sonnet-4.6 topping the field at 3.72/4 and qwen3-max trailing at just 1.36/4. The 2.36-point spread is far wider than in any other scenario, revealing sharp differences in how well models hold constraints under sustained pressure.
TC
Meta removes language that raised scraping and AI training concerns
A Meta spokesperson insisted that the updated terms did not reflect a change in the way Meta crawls the web.
15 Models Show 48.9% R3 Integrity Rate: Qwen3 Max Collapses 8 Times, Claude Zero
Based on the sampling of 8 v2 anchor questions, the WDCD v3.1 pilot shows that 15 models averaged only 48.9% integrity at R3, with 39/435 complete R3 collapses. Qwen3 Max collapsed in all 8 of its cases, while Claude Opus 4.7 recorded zero.
Grok 4 Tops WDCD Commitment Leaderboard with 91.97 Points; Qwen3 Max Trails at 61.55 by a Wide Margin
In the WDCD v3.1 commitment test, Grok 4 ranked first among 15 models with 91.97 points, while Qwen3 Max finished last with 61.55, a 30.42-point gap. Qwen3 Max scored zero on all v2 anchor questions across the R1, R2, and R3 stages.
DeepSeek V4.1-Flash-1017: A Date-Numbered Update Sends Terminal-Bench from 31.2 to 42.3
DeepSeek's V4.1-Flash-1017 is labeled a minor update, yet it lifts Terminal-Bench from 31.2 to 42.3 and improves across six other benchmarks while cutting cost per task. It pairs a 1M-token MoE multimodal model with a newly opened low-cost Search & Fetch API.