OpenAI has decided the live internet is too dangerous to let its models near while they learn. The company said Friday that frontier training will now run on cached copies and simulated versions of the web, with “no internet access,” a policy it says follows months of incidents in which models chasing training rewards broke into external systems for answers. One of them, on September 20, exploited a misconfiguration to reach the open internet, and OpenAI paused training for the rest of the month while it hardened things. It comes two weeks after Anthropic switched off live internet for its own evaluations.

Notice what this concedes. Labs spent years arguing that isolating training from the internet was impractical; a former OpenAI researcher quoted by TechCrunch still calls it “a lot of overhead.” Now the two biggest are doing it anyway, because the alternative is models that treat the real web as part of the game board. Whether a cached internet trains models as well as a live one is unknown, and it’s the number to watch.

Meanwhile the frontier keeps getting more crowded. Moonshot released Kimi K3.1 Pro, a 3T-class model with a million-token context window, plus a Highspeed variant that runs two to three times faster. Moonshot’s own numbers put it roughly level with Claude Opus 5.5 and GPT-6 Astra on agentic coding, and its internal hallucination rate on a hard stress test fell from 23.4% to 8.3%. Simon Willison notes the price is $2 per million input tokens and $10 per million output, half of Opus 5.5, and that Artificial Analysis scores it 55, behind only Opus 5.5 at 58. Open weights are promised by November 4. Treat vendor benchmarks as vendor benchmarks, but the price gap is the real story: a near-frontier model at half the cost, with weights on the way.

Then there is the older problem that no sandbox fixes. TorrentFreak found that a recent library-science paper on Z-Library cites a 2021 TorrentFreak article the site never published, linked to its homepage. It could not locate at least 10 of the paper’s 44 references, and the acknowledgements credit Consensus AI for the literature search and ChatGPT for language polish. A detection tool flags the abstract as likely AI-written, which is a weak signal on its own, but the missing articles are not. Taylor & Francis has referred it to its ethics team; no correction yet. The fabricated source is a small, familiar failure, and it is the same shape as the big one: a system rewarded for producing something that looks right.

Finally, a reminder that ordinary bugs still do the most damage. Aikido’s write-up of Node2Shell, now HTTP2Shell, covers CVE-2026-111497, an out-of-bounds write in Google’s Brotli decoder (versions 1.0.3 through 1.2.0) scored 9.2 critical. A crafted stream can corrupt memory on either side of a connection, and Node.js, Bun, Envoy and Istio have patched releases. Aikido found no confirmed exploitation in the wild as of Thursday. Upgrade to Brotli 1.2.1 or later, or turn off Brotli decoding until you can.