OpenAI pumps the brakes on its next model over cyber capabilities
OpenAI says internal evaluations of Astra, an unreleased model, look strong enough that it can no longer rule out the “Critical” cybersecurity level in its Preparedness Framework. That level describes a model that can find and build working zero-day exploits (flaws the vendor doesn’t know about and has no patch for) against hardened real-world systems without human help, or plan and execute a novel attack end to end from a one-line goal. Testing continues and OpenAI has not confirmed the threshold was crossed, but it acted first: tighter internal controls, blanket monitoring of the model’s agentic use, a pause on internal Astra work that doesn’t yet meet the strengthened security requirements, and outside testing with government agencies and safety organizations. A safety framework only means something if it can make its owner slow down; this is a checkable case of that happening.
Anthropic loosens Fable 5’s biology safeguards
Anthropic rewrote the rules its biology safety classifier follows, brought in outside experts to review them, and cut false positives sharply: biology-related fallbacks (where a request gets rerouted to a weaker model) dropped about 85%. Health questions, biology education, and clinical tasks now go through; dual-use requests in virology, toxicology, and molecular design still get blocked. Over-blocking is a real safety cost too, because it pushes legitimate researchers toward models with weaker protections. Drawing the boundary accurately is much harder than simply tightening it, and that is what this update works on.
OpenJDK, under Oracle, bans AI-generated content outright
OpenJDK’s interim policy bars contributions containing anything generated in part or in full by large language models or similar systems. That covers source code plus docs, pull requests, emails, wiki pages, and bug reports; using AI privately to understand, debug, or review code is fine. Two reasons: the Oracle Contributor Agreement requires contributors to own the rights to what they submit, and the legal status of AI output is unsettled; and plausible-looking but wrong code raises review load on a codebase that critical business systems depend on. The Governing Board approved the policy in late March and published it in April (see the board announcement), so this is not new this week; it is back in the news because Larry Ellison is betting Oracle’s future on AI and declaring that “the code that Oracle is writing, Oracle isn’t writing. Our AI models are writing” (see The Register). Meanwhile GraalVM, also Oracle’s, allows AI contributions, so even a single company has no unified answer yet (see InfoQ).
Cloudflare ships Kitesurf, a browser built for agents, not people
Cloudflare released Kitesurf, a browser engine that runs in V8 isolates on Workers (lightweight JavaScript sandboxes, so there is no full browser process per session). It drops tabs, extensions, and pixel-perfect rendering, and optimizes for what agents actually do: fetch pages, take screenshots, extract content. Cloudflare’s benchmarks show 3–7× less CPU and memory than Chromium, at the price of roughly 1.7× slower wall time and imperfect CSS rendering; the beta is free and works with Puppeteer, Playwright, and MCP clients. Once agents are the main “user,” the browser gets rebuilt around tokens and cost instead of rendering fidelity, and that rebuild has now started.
Rippling burned through its AI budget, then built a per-employee ROI tool
TechCrunch reports that by March, Rippling’s token spending was on track to eat 40% of its R&D headcount budget, growing 80% month over month: one engineer was spending $50,000 a month, and 10–15% of employees drove 60% of the total. The company built AI Spend Console, which tracks spend and output per person and per team, and routes prompts through a gateway to cheaper models that are good enough for the task. July’s token volume roughly matched the April peak (about 600 billion tokens) at 37% of the cost, so the savings came from routing, not from using less. The tool is now for sale (TechCrunch reported in June that about 560 companies already use it), which tells you AI spend governance is turning into a product category of its own.
What the EU AI Act requires of “AI therapy”
artificialintelligenceact.eu, a site that tracks the Act, maps out what applies when general-purpose AI is used for therapy or emotional support. Systems that exploit vulnerable users and cause significant harm fall under the Article 5 prohibition, which turns on a system’s objective or its effect, so a provider that never intended the harm is still covered; systems that count as medical devices requiring third-party conformity assessment are treated as high-risk; Article 50 requires telling users they are talking to an AI; and model providers must assess systemic risks to public mental health and report incidents causing serious physical or mental harm. Enforcement powers took effect on August 2, 2026. The hard part is the line itself: the Act gives no clean test separating a system that manages a mental health condition from one sold for “well-being.”
US states use AI to clear decades of regulatory backlog
Stanford HAI reports on states running AI over statutes and reporting requirements: 500 million words of state code scanned across all 50 states. Maryland found that 20% of reports across four agencies could be cut or merged, including one that took 3,500 staff hours and over $870,000 to produce; New York’s governor ordered a “regulatory reset”; San Francisco passed legislation trimming over a third of its reporting requirements. The error cost here is contained: in each of these cases the AI surfaced candidates, and the cuts themselves ran through the usual channels — Maryland’s 20% are candidates for elimination, not eliminations, and San Francisco’s trims had to pass as legislation. That division of labor, AI nominates and humans decide, is worth copying elsewhere in the public sector.
Today in one line: OpenAI paused parts of Astra’s development because it couldn’t rule out a dangerous capability; Anthropic loosened Fable 5 because it was blocking too much. A safety process that works can do both: hit the brakes, and admit its own false positives.