Ambient Advantage

Anthropic discovers a hidden cognitive workspace inside Claude that reveals what the model thinks but never says. Plus: the first fully autonomous AI ransomware attack hits production, OpenAI's GPT-5.6 goes public, and Meta's image generator sparks a privacy backlash.

The first documented agentic ransomware attack hit production systems with no human at the keyboard. Today we cover JADEPUFFER, Anthropic's J-space interpretability breakthrough, DeepSeek designing its own chip, and why the agentic threat landscape just escalated on two fronts simultaneously.

Anthropic discovers Claude has a hidden internal workspace called J-Space that could transform AI safety monitoring. Plus: the first fully autonomous AI ransomware attack, OpenAI's government-gated GPT-5.6, and Nvidia's next-gen rack delayed to 2028.

Claude Fable 5 now completes 16% of real freelance projects at professional standard, quadrupling the frontier in just eight months. Plus: Sonnet 5 makes agentic AI the default tier, Tesla and Uber slam the brakes on token spending, and Altman proposes an IAEA for AI.

Claude Fable 5 is back online after an 18-day government-imposed freeze, setting a precedent that frontier models can vanish on 90 minutes' notice. Plus: Sonnet 5 launches at half the cost of Opus, GPT-5.6 remains in government review limbo, Meta enters the cloud wars, and Etched emerges with $1B in AI chip contracts.

OpenAI's GPT-5.6 launches under government-gated access while Anthropic ships Sonnet 5 and brings Fable 5 back online. Plus: Etched challenges Nvidia with a $5B inference chip, Ford's cautionary AI tale, and why access and governance now matter more than benchmarks.

Anthropic makes Claude Sonnet 5 the default across all plans with near-flagship agentic performance at 40% lower cost. Plus: OpenAI's GPT-5.6 launches behind a government gate, Etched exits stealth with a billion dollars in chip orders, and the AI market splits in two.

OpenAI launched GPT-5.6 but only 20 government-vetted partners got access, signaling a new era of regulated frontier AI. Also: Google rations Meta's compute, OpenAI unveils its custom Jalapeño chip, Five Eyes warns AI cyberattacks are months away, and Austria lobbies the EU to host Anthropic.

Washington took the wheel on frontier AI this week. OpenAI launched GPT-5.6 under government-vetted access restrictions while Anthropic's Claude Fable stays dark on day 17. Also: Five Eyes warn AI cyberattacks are months away, a Chinese open-source model matches the banned capabilities, and the AI sovereignty scramble hits Europe.

The White House restricts GPT-5.6 in the first-ever pre-release gate on a U.S. AI model, while Anthropic accuses Alibaba of a 28.8 million interaction distillation attack on Claude. Plus: OpenAI's custom Jalapeño chip, Claude lands in Slack, Gemini undercuts on agentic pricing, and AGI timeline claims converge.

OpenAI and Broadcom unveil Jalapeño, a custom inference chip built in just nine months with AI-assisted design. Plus: Anthropic embeds Claude directly into Slack, Cursor launches an agent-native GitHub rival, and Five Eyes warns AI cyberattacks are months away.

Anthropic's Fable 5 and Mythos 5 remain offline after the first-ever government-forced model takedown, and the Five Eyes alliance warns AI cyberattacks are months away. Plus: Sakana's multi-model hedge, SpaceX's $80B compute empire, and AI that out-persuades world champion debaters.

Five Eyes intelligence chiefs warn AI-powered cyberattacks are months away, urging boards to treat cyber risk as a leadership responsibility. Plus: OpenAI expands Daybreak with 30 security partners, a Nobel laureate defects from DeepMind to Anthropic, and a landmark study proves AI now out-persuades expert humans.

Google DeepMind now treats its own AI agents as insider threats, publishing a security framework modeled on MITRE ATT&CK. Today's briefing also covers Builder.ai's $1.5B fraud collapse, self-improving agent memory from Perplexity, OpenAI's confidential IPO filing, and why 13 words can poison AI search.

The Pentagon confirmed Grok AI helped target 2,000 strikes in Iran, marking the first commercial LLM used in live combat. Also: SpaceX buys Cursor for $60 billion, Noam Shazeer leaves Google for OpenAI, ChatGPT drops below 50% market share, and Adobe launches its agentic creative workflow.

SpaceX acquires Cursor for $60B, signaling the AI competition has shifted from models to distribution and control. Today's briefing also covers Microsoft Copilot Cowork going live, Anthropic's Fable 5 getting pulled by the U.S. government, DeepSeek's $7.4B raise, and the enterprise AI cost crisis.

The US government suspended Anthropic's most capable models days after launch, creating a new category of sovereign risk for every enterprise AI deployment. Also covered: a German court holds Google liable for AI hallucinations, Uber blows its AI budget in four months, and critical security flaws hit LangGraph and Splunk.

The US government pulled Anthropic's Fable 5 offline via export controls — the first time a commercial AI model has been killed by regulatory action. Plus OpenAI files for a trillion-dollar IPO, agentic AI budgets explode 6x, and Google cuts AI subscriptions to five dollars a month.

Anthropic's Fable 5 is the most powerful public AI model yet, but its system card reveals it silently throttles what it helps you build. We cover the hidden guardrails, Amodei's sweeping regulatory proposal, the OpenAI-Anthropic price war, Uber's budget blowout, and dual IPO filings reshaping enterprise AI.

Anthropic launches Claude Fable 5, the first Mythos-class model available to everyone, with a mandatory data-retention policy that changes the enterprise compliance picture overnight. Plus: OpenAI files for its IPO, Apple rebuilds Siri on Google Gemini, and Uber burns through a $3.4B AI budget in four months.

Anthropic launches Claude Fable 5, a new top-tier model with a dual-version safety architecture. OpenAI files its S-1 at $852B while declaring chat dead, Uber blows its AI budget in four months, and the Miasma worm turns AI coding agents into attack vectors.

OpenAI and Anthropic both file for IPOs, forcing frontier AI labs to open their books for the first time. Plus: Apple bets Siri's future on Google Gemini, Uber blows its AI budget in one quarter, and Anthropic drops a zero-trust security blueprint for AI agents.

Anthropic calls for a global AI development pause, revealing that 80% of its code is now written by Claude. Plus: Google signs a $920M/month compute deal with SpaceX, a Canadian research team demonstrates an AI worm that adapts its attacks per target, and ChatGPT launches self-updating memory.

Researchers hijacked Google Gemini through a WhatsApp message, and Meta's AI chatbot handed over Instagram passwords to hackers — two incidents that expose a structural flaw in how AI agents handle trust. Plus: DeepSeek raises $7.4B, Anthropic files for IPO, Microsoft unveils its own models, and Canada launches a $200B national AI strategy.

Microsoft Build unveiled a full agentic AI stack with seven in-house models, while Anthropic filed its S-1 and Nvidia reimagined the PC as a local agent machine. We also cover OpenAI Codex going white-collar, the open research brain drain, Trump's new AI executive order, and why GitHub Copilot pricing is burning developer budgets overnight.