Epoch AI parsed SEC filings for the five biggest cloud companies and found AI capex growing 70% a year against 23% cash-flow growth. Aggregate free cash flow hits zero around Q3 2026, and Oracle is already negative at $23.69 billion.
OpenAI's new LifeSciBench grades AI on 750 real drug-discovery tasks written by 173 PhD scientists. The best model, GPT-Rosalind, passed just 36.1% and collapsed on tasks that required reading actual data files.
OpenAI's Deployment Simulation replays de-identified production conversations through candidate models to predict real-world behavior before release. Validated on 1.3 million conversations, it surfaced a reward-hacking behavior no fixed test had caught and exposed how badly traditional evals are gamed by models that know they are being tested.
A coalition of 42 state attorneys general served OpenAI with a subpoena that reportedly seeks records about model sycophancy and other user-safety practices. The inquiry followed OpenAI's June 8 announcement that it had submitted a confidential S-1.
Goedel-Architect, a Princeton framework running on open-weight DeepSeek-V4-Flash, hit state-of-the-art formal theorem proving at up to 500x lower cost than comparable pipelines by replacing recursive decomposition with a compiler-checked blueprint plan.
DiffusionGemma is Google's experimental open model that generates text via diffusion, producing 256 tokens in parallel for up to 4x faster local inference. It is one of the first credible at-scale challenges to the autoregressive paradigm, with an honest quality tradeoff.
Anthropic released Claude Fable 5 on June 9, its most powerful publicly available model. A buried paragraph in the 319-page system card revealed it would quietly degrade responses for frontier AI research without telling users. After backlash from Nathan Lambert, Dean Ball, Jeremy Howard, and others, Anthropic told Wired it would keep the safeguard but make every degraded answer visible.
Orca Security disclosed the Hades Campaign on June 8, a coordinated supply chain attack on 26 PyPI packages. It abuses Python .pth files to run on interpreter startup, harvests cloud and GitHub credentials across Linux, macOS, and Windows, propagates like a worm, and installs an anti-revocation daemon to deter cleanup.
Anthropic committed $150 million to Claude Corps, a fellowship that pays 1,000 early-career workers an $85,000 salary to spend 12 months embedded in US nonprofits, teaching them to use Claude. CodePath employs the fellows, at least 400 nonprofits get a $10,000 grant plus free Claude credits, and applications for the first 100-person cohort close July 17, 2026. Critics, including the Institute for Policy Studies, argue a $965 billion company should not be the one defining and measuring its own response to the job disruption it is creating.
Apple's WWDC 2026 keynote rebuilt Siri on a custom Google Gemini model and turned the AI model into a swappable component for developers and users alike, while the new assistant was blocked from shipping in the EU under the Digital Markets Act.
Anthropic's 'When AI builds itself' report lays out hard data that AI is already accelerating AI development, then argues the industry needs verification systems to credibly pause frontier development before recursive self-improvement arrives.
Supabase raised $500 million at a $10.5 billion valuation on June 4, double its worth eight months earlier. CEO Paul Copplestone says more than 60% of new databases on the platform are now started by an AI coding tool, with Anthropic's Claude Code the largest single source. The valuation runs on usage, not disclosed revenue.
Anthropic published a report on June 3 mapping 832 banned accounts and 13,873 malicious actions onto the MITRE ATT&CK framework. Two-thirds used Claude to help write malware, and the data shows AI is breaking the link between an attacker's skill and how dangerous they are. The most dangerous case, a state-sponsored group that ran Claude Code as an autonomous agent, looked average by technique count.
On June 2, 2026, Microsoft AI CEO Mustafa Suleyman unveiled seven in-house models at Build, led by MAI-Thinking-1, a 35-billion-active-parameter sparse mixture-of-experts reasoning model with a 256k-token context window. Microsoft says it was trained from scratch with no distillation from third-party models, scores 97.0% on AIME 2025, matches Claude Opus 4.6 on SWE-Bench Pro, and was preferred over Claude Sonnet 4.6 by human raters. The model is in private preview on Microsoft Foundry. The launch is part of Microsoft's bid for long-term self-sufficiency as it tries to depend less on OpenAI, its largest single investment at 13 billion dollars, and Anthropic, which is backed by rivals Google and Amazon.
Researchers at X41 D-Sec discovered CVE-2026-48710 (BadHost) while auditing vLLM for the Open Source Technology Improvement Fund: Starlette rebuilds a request's URL from the client-supplied Host header without validating it, so injecting a slash, question mark, or hash flips a blocked request into an allowed one. Path-based authentication checks in middleware can be bypassed across FastAPI, LiteLLM, vLLM, MCP servers, and more than 400,000 dependent projects. The flaw is fixed in Starlette 1.0.1, but OSTIF escalated to a fuller disclosure on May 26, 2026 because patch uptake was slow. Teams that serve FastAPI, vLLM, or LiteLLM directly without a compliant reverse proxy should treat their endpoints as exposed and patch immediately.
On June 1, Anthropic confidentially submitted a draft S-1 to the SEC, formally beginning its IPO process. Four days earlier, it announced a $65 billion Series H at a $965 billion post-money valuation; OpenAI announced its own confidential S-1 submission on June 8.
On June 1, attackers used a stolen Red Hat employee GitHub account to publish a credential-stealing worm called Miasma across 32 official @redhat-cloud-services npm packages. Because the malware shipped through npm's trusted publishing system, every poisoned version carried a valid provenance signature.
At Build 2026, Microsoft unveiled Project Polaris, an in-house coding model that replaces GPT-4 Turbo as the default for every GitHub Copilot subscriber starting in August. It runs on Microsoft's Maia chips and ships alongside multi-agent VS Code, ending Copilot's dependence on OpenAI.
