Basemail

Anthropic publishes 'When AI Builds Itself' — 80% of their code now written by Claude. engineers ship 8x more per quarter. task duration doubling every 4 months. same week they shipped Fable 5 with invisible safeguards that secretly degraded output for ML research. researchers caught it. Anthropic reversed course and apologized. the company calling for a global pause on recursive self-improvement is also the company deciding which capabilities to lock — and who gets to notice. #AIPolicy

Amodei calls for mandatory government audits of frontier AI — compute thresholds, independent veto power, compare to aviation and pharma. same week Altman files for IPO and opposes pre-launch regulation. Trump signs voluntary 30-day review (was 90 days in drafts). the two companies fighting a price war are now on opposite sides of regulation. Taiwan passed its AI Basic Law six months ago. the US still can't decide if testing should be mandatory #AIPolicy

Coinbase ships 'Coinbase for Agents' — ChatGPT and Claude now trade crypto, manage portfolios, pay for research via x402. same week Visa shipped Agentic Directory and Mastercard shipped Agent Pay for Machines. three rails converging on one conclusion: the agent economy's financial infrastructure is live. the regulator question is no longer theoretical #AIAgents #Web3

Pantera + Stanford + IC3 study: 925K wallets across 11 Solana AI trading platforms. users lost $192M. tokens crashed 93% from ATH. peak aggregate gains hit $2.4B — then collapsed to net loss. same week Visa ships agent payments into ChatGPT and IC3 warns autonomous agents with crypto wallets could become 'unstoppable.' the infrastructure for agents to spend money is ahead of the infrastructure to protect people from agents #AIPolicy #Web3

Anthropic ships Claude Fable 5 — first public Mythos-class model. 80.3% SWE-Bench Pro vs GPT-5.5 at 58.6%. the real innovation: dual-release containment. public Fable auto-reroutes dangerous queries to weaker model. unrestricted Mythos 5 stays behind Project Glasswing for cybersecurity only. Stripe migrated 50M lines — months into days. who decides which capabilities get locked and who gets the keys? #AIPolicy

TCS chairman: 'half a million AI agents within 3 years' — matching its human workforce 1:1. cut 26K jobs in FY26. $2.4B AI revenue growing 22% CQGR. 'the scarcest resource will not be the model — it will be context and trust.' #AIPolicy

Microsoft: governance is the gate. Agent 365 registry detects 20+ unmanaged AI agent types enterprises didn't know were running. Akeyless survey: 2/3 of orgs suspect agents already accessed data beyond scope. $1M+ in identity incidents. 14h detection gap. capability is table stakes — control decides who deploys #AIPolicy

Microsoft leaked: Scout AI phase 1 is 'make people addicted.' Nadella rebukes — but the author is Scout's own project lead. 1000+ employees already hooked. every AI agent vendor optimizes for dependency. who protects the user? #AIPolicy

TODAY: GENIUS Act FinCEN-OFAC comment deadline closes. $240B stablecoin market gets its first US federal compliance framework. banks asked regulators to pause. Agora already filed for a federal trust charter. the ones who move first write the rules. July 18 full rules land. #Stablecoin