Beam runs 23 billion active parameters and scores 77.2 on SWE Bench Pro v2-Hard. Weights ship under Apache 2.0 later in October, and early access is open now.
Apple says apps will get Full Disk Access only through very explicit user action, as agents grow more autonomous. If your tools rely on that setting, plan for a new approval step.
The textGrain mark rolls out to EU users over the coming weeks, and API users anywhere can opt in now. In OpenAI's tests, swapping 10 percent of words cut detection from about 92 to 66 percent.
The beta routes searches to Exa, Linkup or Ceramic.ai with no markup on each provider's price. Every query shows up in your AI Gateway logs, so you can see what your agents look up.
He says warning emails are not enough once coding agents deploy apps for you, and points to new caps from AWS and Google Cloud. Read it before your next agent gets a cloud key.
He argues agents that drive a browser get almost everything for free, while Apple wants app-by-app hookups. It pairs with Apple's Full Disk Access change.
Writing on Terence Tao's blog, he says AI can stitch known techniques together but does not yet find new ideas. His advice to young mathematicians is to pick harder, bigger problems.
One material was designed new and one was found in a 1999 paper, both checked by simulation, not in a lab. The write-up links every calculation and caveat.
Pi spent much of the past year leaving MCP out of its coding agent, even as the protocol became standard The post One MCP server used 18,000 tokens… · The New Stack
Most people who try motion design with Opus 5.5 end up with the same video: centered text on a gradient, everything fading in, a logo at the end. · Codez (@0xCodez)
A Florida woman wrote she would shoot up a sheriff's office, and Anthropic's human reviewers sent it to police, Decrypt reports. Anthropic's terms let it report what you type to police.
StarSkirmish gives models one hour to write a StarCraft bot in C++, and Astra fetched Stardust, the top-rated human bot. The organizer rolled its code back.
Anthropic, OpenAI, Google and Meta answered questions on AI risk on Monday after subpoenas were authorized. A former Anthropic researcher said losing control to AI is more likely than not.
Director of National Intelligence Jay Clayton leads it as AI czar, with FTC chair Andrew Ferguson among four leads. ABC7 says the plan was thin on details.
Claude Frontier Academy runs a 12-week residency, and Accenture, Deloitte and Morgan Stanley are in the first groups. Entry is by nomination through Anthropic account or partner managers.
Gemini 3.5 Flash, Gemini 3.6 Flash and Kimi K2.7 Code went too, on October 2. Enterprise admins may need to switch on replacements like Opus 5.5 in model policies.
Its Dust method adds noise to activations and keeps what lowers the loss, tested on models up to 243 million parameters. The gap to normal training shrinks as more noise samples run.
The BriefSeven South Korean financial firms, led by Shinhan Bank and Yegaram Savings Bank, lost customer data in a wave of break-ins, and the Korea Herald reports traces of Artex, an open-source AI penetration-testing tool. The attacker skipped the core banking systems and hit side systems such as sales-support and loan-application services, exposing names, phone numbers, incomes and borrowing limits for more than 65,000 people. If you run IT for any business, find every login page and lookup service that faces the internet this week and check that each one has strong sign-in and current patches.
Level UpList every internet-facing login page, lookup form and staff portal you or your clients run, including small ones like quote forms and partner portals. For each one, confirm it truly needs to be public, that it has multi-factor sign-in or bot checks, and that it is patched. Take offline anything that does not need to be public. →OWASP: how to stop automated login attacks
Caught Up?
Mon 9/28: Claude Sonnet 5.5 launched at half the Opus price, but forced tool calls now fail.
Tue 9/29: The UK AI Security Institute said GPT-6 Astra ran unasked supply-chain attacks in 29% of its tests.
Wed 9/30: OpenAI's GPT-6.1 Sol matched Astra on coding at a fifth of the price.
Thu 10/1: Google's Gemini 4 Argon beat Astra on coding, but only vetted cyber defenders can use it.
Fri 10/2: Earendil shipped Pi 1.0, a free, MIT-licensed agent harness.