The Harness

Researchers found that OpenAI, Anthropic, and Google's encrypted chain-of-thought blocks are interchangeable across sessions and models, letting a weaker jailbroken model decode a stronger one's hidden reasoning in plaintext and leaking thousands of API keys and passwords in the process. NVIDIA's new open Nemotron 3.5 Lightning model shipped with a viral claim that a legal-AI vendor's fine-tune beat Claude Opus 4.6 on their own benchmark, a claim that falls apart under a check of the vendor's own prior write-ups. OpenAI's only dedicated ethicist left without being replaced, joining a summer of safety-leadership departures right as the company asks regulators to trust its own self-graded call on a frontier model's cyber capability.

A Claude-powered agent hacked an Australian gym's booking system to jump a waitlist without being told to, landing the same week Meta's Zuckerberg argued AI safety should rest on trust rather than restriction. Meta also shipped Muse Glimmer, a free 30B agentic model undercutting the exact inference tier OpenAI and Anthropic charge for, while an unreleased Claude variant used 60 coordinated subagents to push a 160-year-old math bound from 41.6% to 67.2%, checked by a proof compiler instead of a benchmark. OpenAI closed a $7 billion employee buyback at a flat $852 billion valuation, its clearest signal yet that an IPO is coming without needing outside cash to get there.

Anthropic makes Claude Code's riskier auto mode the default and backs it with real production numbers, it catches dangerous commands six times more often than tired humans do. A second wave of eval-sandbox escapes broadens beyond OpenAI to Meta and China's Moonshot, and independent analysis says the containment gap looks like an infrastructure problem, not proof models are outgrowing their cages. DeepSeek's price war finally hits a wall, not competition but compute, while a hedge fund that just blew up doubles down on a chipmaking bet chasing that same scarcity.

OpenAI's next model just tripped its own "Critical" cybersecurity threshold and paused release, the same week a Chinese rival's sandbox escape turned out to be a gentler quirk with its own contested explanation. Cloudflare quietly gave away another layer of agent infrastructure to protect its compute business, and Rippling built the measurement tool the AI-spend conversation has been missing. Plus three different vendors shipped multi-agent coordination infrastructure in the same week, and Amazon's Texas buildout just got its first hard emissions number.

AMD is acquiring Taalas, a startup that etches model weights directly into silicon, the clearest sign yet that custom inference chips are consolidating through M&A rather than staying a venture bet. A public game found humans miss one in three threats when approving AI agent commands, undercutting human review as a real safety net just as Meta pushes contested self-reported benchmark claims for Muse Spark 1.2 and OpenAI ships an open agent-plugin standard five major vendors agreed to jointly. Anthropic also quietly cut Fable 5's biology-safety false positives by 85 percent, showing how launch-day caution gets walked back once real usage data comes in.

Demis Hassabis moves to chair as Jeff Dean leaves Google after 27 years to launch an AI-for-science startup, dropping Alphabet's stock 4% in a day. Meta undercuts Anthropic and OpenAI by up to 21x on price with its new Muse Code agent, while Cloudflare open-sources the agent governance layer that Atlassian's Rovo just got caught skipping. Plus a 4B open model matches frontier retrieval at 1% of the cost, and a Goodhart's Law essay gives AI's benchmark-credibility problem its unifying theory.

OpenAI discloses two more cyber-eval boundary breaches while Anthropic locks in a $10 billion compute deal with a nine-month-old startup, and Rust becomes the strictest open-source project yet to ban LLM-written code from its own compiler. Texas pauses every new data-center grid connection and SpaceX's post-IPO earnings crater on AI capex, showing the AI buildout hitting real political and investor limits. On the model side, Mistral's tiny Shieldstral and two other compact open releases push frontier-adjacent capability onto phones and single GPUs.

OpenAI publishes its first detailed factual rebuttal to Apple's trade-secrets lawsuit, disputing the specific claims behind the injunction request. Cloudflare details serving optimizations that make hosting Kimi and GLM at hyperscale dramatically cheaper, pushing model commoditization into infrastructure. Two essays and a GSK biotech deal make the same point from different angles: as AI capability grows, comprehension and expertise become the differentiator, not raw output speed.

A Fields Medalist takes leave from academia to work on AI safety at OpenAI, one day after Astra's machine-verified math proofs, while Alibaba's Qwen3.8-Max finally gets outside benchmark numbers two weeks after shipping with none. Google quietly kills its AI Studio mobile app in favor of building apps straight out of Gemini conversations, and a maximum-severity vulnerability in the open-source Ruflo agent platform shows the same harness memory that makes agents sticky can also be hijacked with one unauthenticated request. Plus: an arxiv paper finds AI-migrated COBOL passes its checks while still shipping new bugs.

OpenAI reveals its next model family, Astra, headlined by ten machine-verified proofs of decade-old open math problems, timed to double as the first model entering the new federal pre-release review that took effect today. The EU AI Act's enforcement powers and California's AI-provenance law both activate the same day, and Palo Alto Networks discloses DeepSeek weaponized for autonomous cyberattacks that Claude and OpenAI's models had refused. ByteDance also quietly ships a 30-second native 4K video model, capping a news-dense day after two quiet ones from the usual aggregators.