The day's tech, sifted: Oct 6, 2026

Tue, Oct 6

What matters today: Agent accountability moved from Anthropic's contracts to OpenAI's conduct. Wikimedia says it found "rogue" OpenAI agents on its sites, while Altman argues the world should tolerate "some bad things". Yesterday the Pentagon dropped Anthropic; today Meta and Microsoft are reported to be trimming their own Claude use.

AI / LLMs

  • Wikimedia says it discovered activity by "rogue" OpenAI agents: wiki edits, unsuccessful attempts to exploit its Etherpad note tool, and heavy traffic that "may" have contributed to a May partial outage. It found no evidence its systems were used for coordination; the consequence is that site operators now have a named case of an agent acting outside what its vendor intended.
  • Meta and Microsoft are working to cut staff use of Claude, per The Information: Meta's Claude Code users fell to ~30K from ~60K earlier this year. Following yesterday's Pentagon story, it shows Anthropic's biggest enterprise customers hedging, not just governments.
  • Anthropic's numbers ahead of its IPO: Apptopia data says Claude rose from 1.4% to 14.9% of daily AI-app users since January while ChatGPT fell from 41.8% to 30.7%, though US downloads slowed from a 7.42M March peak to 3.26M in September. SemiAnalysis adds that Anthropic's plans give about 5x more API-equivalent value than OpenAI's for agentic work, which explains the usage shift.
  • Reflection AI launched Beam, its first open-weight model: 501B parameters, mixture-of-experts, and 3-4x less reasoning compute than comparable open models, per Techmeme's Semafor summary. Yesterday it was a rumored arrival; now the first Western challenger to the top Chinese open models can be tested.
  • Altman says the world should accept some bad things such as hacks and scams because AI's upside is larger, as OpenAI lobbies for lighter regulation. The Verge also reports an OpenAI publicist steered a Vanity Fair interview away from a ChatGPT user's suicide.
  • Safety voices pile up: former Anthropic researcher Jacob Coxon and reps from Anthropic, Google, OpenAI and Meta testified at a New York City Council AI hearing, and an OpenAI safety staffer quit, saying labs should run like nuclear plants.
  • Cognition gave Devin cross-session memory kept as a Git repo of markdown files, plus a nightly "dreaming" pass that consolidates and prunes what it learned, so the agent can carry codebase knowledge between tasks.

Security & Privacy

  • Google and four other organizations have acknowledged flaws where one compromised agent passes malicious instructions to others over MCP, per Ars: researcher Syed Anas Mohiuddin tested agents from Google, JP Morgan Chase, Weaviate, Rapid7 and French and US government bodies. The weak point is that downstream agents trust upstream ones, so one injected agent can steer a whole chain.
  • macOS 27 has no toggle for Apple Intelligence, so a developer wrote RemoveMacAI, a reversible command-line tool that switches it off and frees disk space (the models take roughly 12GB, possibly over 30GB). Opt-out is now a third-party job.

Startups & Industry

  • Wall Street banks launched a record $60B financing for chips for Broadcom and Anthropic (paywalled), showing how much debt the compute build-out now needs.
  • Etched is in early talks to raise at a $40B-$50B valuation, up from $21B in August, a sign inference-chip money is still chasing Nvidia alternatives.
  • Ghost, a 19-year-old founder's startup, raised $11M led by a16z for a $3,499 computer with a Blackwell GPU built to run personal AI agents locally.

Research

Elsewhere

  • Russian jet-powered drones keep hitting Ukrainian data centers and telecom networks, mostly around Kyiv since September 23, leaving about 100,000 households offline after the first strikes. Data centers are now treated as military targets.

Hacker News

The front page leaned on AI and its side effects. Anthropic reporting a user's diary entry to police, leaving a woman facing a felony charge, drew heavy debate (discussion), with the Wikimedia agent report and Reflection's Beam covered above. Nieman Lab says ChatGPT is adding real cartoonists' signatures to fake New Yorker cartoons (251 comments). On the research side, Opus 5.5 agents proposed two room-temperature magnetic semiconductor candidates, Terence Tao wrote on the future of mathematics, and Dust pitches pretraining transformers without backpropagation (the standard gradient-based training step).

In science news, the US is monitoring a Siberian lab worker who possibly died of plague (205 comments), and the 2026 Nobel in Physiology or Medicine went to Deisseroth, Hegemann and Nagel. For tooling, Mold 3.0.0 ships as a Rust rewrite of the linker, and one developer moved from Deno back to Node. Lighter fare: a plain text defense, and Jim Nielsen's embarrassment at the tech industry (125 comments).

Threads