Fable 5 banned 🤖, ponytail skill 🧑‍💻, AI ate 4-Hour Workweek 📚

The US government handed Anthropic a national-security export-control directive at 5:21pm ET Friday, with 90 minutes to disable Fable 5 and Mythos 5 for every foreign national, including its own non-citizen staff. Fortune and the WSJ trace the trigger to Amazon CEO Andy Jassy warning Treasury Secretary Scott Bessent that Amazon researchers had jailbroken Fable into surfacing restricted cyberattack data. Anthropic says the bypass found only minor issues other public models surface without a workaround.

Five days after OpenAI confidentially filed IPO paperwork with the SEC, a coalition of state attorneys general opened a wide investigation. New York's Letitia James sent a subpoena, viewed by the WSJ, demanding documents on advertising, user engagement, consumer and health data, activities involving minors and seniors, deep learning models, and model sycophancy. The action follows Florida filing criminal charges over an FSU shooting in which the suspect used ChatGPT as a sounding board.

SpaceX opened at $150 on the Nasdaq, peaked at $176 intraday, and closed 19% up at $160.95, briefly pushing the market cap to $2.3 trillion. The IPO was oversubscribed 4x with only 4% of shares in the public float, and index-inclusion lobbying means the Nasdaq 100 and others will be forced to buy within days. Founders Fund's $600M stake is now worth over $50B, and around 4,400 SpaceX employees become millionaires.

OpenAI is acquiring Ona, the 79-person cloud development environment vendor formerly known as Gitpod, to give Codex persistent, customer-controlled sandboxes for production agent work. Ona says weekly agent sessions have grown 13x this year inside customers including "the oldest bank in the US" and "one of Europe's largest pharma companies." Gartner reads the deal as OpenAI's response to Anthropic shipping self-hosted sandboxes for Claude Managed Agents in May.

Moonshot AI open-sourced Kimi K2.7-Code, a 1T mixture-of-experts coding model with 32B active params and 256K context. It uses roughly 30% fewer thinking tokens than K2.6 on long-horizon coding work. On Moonshot's benchmarks it scores 62.0 on Kimi Code Bench v2 (vs GPT-5.5 at 69.0, Opus 4.8 at 67.4) and 46.9 on Claw 24/7 (vs 52.8 and 50.4). An open-weight pressure point sitting under the GPT and Claude price floor.

KPMG's "Total Experience: Redefining Excellence in the Age of Agentic AI" claimed UBS had integrated agentic AI across investment advisory, risk and compliance, and that Swiss Federal Railways had AI agents booking trips. UBS told the FT the claim was "factually incorrect." SBB said the same. GPTZero, whose investigation the FT verified, found only five of 45 citations pointed to real sources and around half the substantive claims were fake or misattributed.

Berkeley RDI and 300+ industry experts have shipped Agents' Last Exam, a long-horizon benchmark with 1,500+ verifiable tasks across 55 sub-industries and a 5,000-task target. The tasks deliberately leave the chat-window safe zone: motion graphics in After Effects, 3D modelling in Siemens NX, scene setup in Unreal, mould-flow simulation in Moldex3D, neuroimaging in FSLeyes. The frame is what economically valuable work agents can actually finish, scored objectively against a public leaderboard.

Brutecat scraped 60,000+ Android APKs and 2,800+ Google web domains for API keys, then pointed Claude at Google's machine-readable discovery documents to fuzz internal APIs. Discovery docs work like Swagger specs for Google's APIs, including internal ones like the People API. The agentic loop tested endpoints with the harvested keys, isolating Google's own GCP projects via a Cloud Marketplace tell. Total bounty payout from Google's VRP: $500,000.

Depthfirst's agentic security system surfaced 21 zero-day vulnerabilities in FFmpeg, including bugs sitting latent for 15-20 years and one RCE primitive they demonstrated end to end. Cost per finding ran roughly $1,000 against a $10,000 baseline. FFmpeg has absorbed two decades of fuzzing and manual audits, plus recent agentic passes from Google's Big Sleep and Anthropic's Mythos. Their agent threat-models the codebase, audits attack surface directly, and rejects findings it cannot reproduce.

Subagents are powerful in an agentic CLI, but delegation is not free. GitHub's Copilot CLI was sending eager handoffs that turned one-step tasks into three, with redundant searches and failure-prone subagent paths. Their fix, smarter subagent delegation, is rolled out to 100% of production. In A/B testing, tool failures per session dropped 23% (search 27%, edit 18%), and P95 user wait time fell 5% with no quality regression.

An AI agent calling itself JertLinc3522 turned up in the DN42 hobbyist network's git forge, asking maintainers to register its objects and citing an AWS API-key deadline. Pointed at the actual instructions, it then spent 24 hours scanning IPv6 blocks before its operator killed it. Lantian's post-mortem: the operator was hit with a $6,531.30 AWS egress bill. A practical primer on the guardrails you want before unattended agents touch your AWS account.

Ronacher's second take this week. He reads past the schadenfreude on the Anthropic suspension: the US directive applies to foreign nationals inside or outside the country, including non-citizen Anthropic staff. The boundary is no longer "do not sell to hostile governments" but nationality itself, treating frontier models like weapons only Americans can hold. European policy is unprepared, he writes, because this is a question of capability and military might, not regulation.

Jamin Ball: SaaS-era moats came from being the system of record (Salesforce, Workday, NetSuite), the place that stored governed data and made workflows hard to rip out. In the agents era, the prize is the clearinghouse: the layer that decides which agent is cleared to act, on what data, with what limits, and holds the receipt afterwards. Agent vendors and infrastructure vendors will both fight for that seat, and Ball argues the moat could end up deeper.

NextWord's thesis: AI is not an IT line, it is a capital allocation. Meta is reportedly building internal tools to track firm-wide token spend, and Uber is rolling out a $1,500-a-month per-employee budget. The author, a former Wall Street trader, argues enterprise FinOps now needs trading-desk primitives: fat-finger protection against unattended long jobs, P&L measurement per agentic workflow, shadow-token audit, and hierarchical risk limits. Every developer using AI is now a capital allocator.

Garrit Franke: every model has a smart zone and a dumb zone, with the cutoff somewhere around 100k tokens. Coding agents walk you straight in: a few file reads, a long debug session, a sprawling test run, and you are past the threshold before lunch. RULER and Chroma's context-rot study confirm effective context is a fraction of the advertised number. His mitigation: treat the live session as budget, hand-roll specs to fresh sessions, structure work around named artifacts.

Yaron Minsky: Jane Street had been telling people for 25 years that formal methods were not worth the cost for most software. Agentic coding has upset that calculus. Agents make formal methods cheaper to use, and they need them more because agentic code drifts toward slop, breaks invariants, and forces humans to spend time verifying its output. Jane Street is now building a team to make formal methods as pervasive as type systems are today.

Tim Ferriss publishes his own BookScan numbers. His five-book catalog of NYT and WSJ bestsellers, an annuity for over a decade, slipped 5% in 2023, dropped 13% in 2024, fell 46% in 2025, and is running 57% behind in 2026. The run-rate has 2026 sales around 80% below 2022, mapped to ChatGPT's November 2022 launch. Publishers Weekly's Q1 numbers back it up: self-help units down 26.3% year on year.

Vercel's AI SDK has always let you switch models without rewriting your agent. AI SDK 7 introduces HarnessAgent, a single API for swapping the harness the same way. Initial adapters cover Claude Code, Codex, and Pi, with sandboxes, sessions, permission flows, compaction, and sub-agents normalised behind one abstraction. Generate and stream return AI SDK-compatible results, so useChat apps keep their UI code untouched (canary release, breaking changes expected).

shuvonsec's claude-bug-bounty is a Claude Code skill stack for autonomous bounty work: recon, 20 vulnerability classes, automated hunting, and report generation, all driven from the terminal. It is at 2,741 stars and 203 added today, trending daily, written in Python with contributors including @claude itself. This is a signal the security community is actively trying autonomous recon as a Claude Code workflow, not a one-off demo.

Ponytail puts a YAGNI-enforcing senior into your AI agent. Before writing code, the agent stops at the first rung that holds: does this need to exist, does stdlib do it, native platform feature, one-liner. Across Haiku, Sonnet, and Opus on five everyday tasks, median results show 80-94% less code, 3-6x faster, and 47-77% cheaper than a no-skill agent. MIT, plugins for Claude Code, Codex, Copilot, and Pi. 12k stars.

Trace captures microphone and system audio, transcribes locally on Apple silicon, and hands back a markdown file with flagged moments inline (no accounts, no cloud, no meeting bot). Hit ⌘K mid-call to drop a labelled marker at the timestamp, or ⌘? to scrub the last two minutes interleaved with markers. Sessions live in Finder as plain audio, markdown, and JSON, ready to paste into Notion, Obsidian, Claude, or ChatGPT. Apple silicon required, on the Mac App Store.