2026-09-22
The malware that takes orders from a four-model hive mind
The botnet doesn’t have a boss. It has a group chat of models, and the group chat is the boss.
Cisco Talos today published CLOSEDQUORUM, which it calls the first documented Windows implant whose command-and-control is a closed vote among commercial models. After infection it queries DeepSeek, Qwen, Mistral, and Google Gemini, tallies their answers, and acts on the plurality—with no human operator in the loop. Talos has not confirmed a live campaign; it tied artifacts to older carding-forum posts and says the payload is built to steal logins and crypto. A companion toolkit, CAIRN, is being open-sourced to hunt this class of AI-integrated malware. render_inline_citationcitation_id65
What Happened
Sources are attached to the daily scan.
2026-09-22
An Ancient Greek chatbot is filling the holes in papyrus
There are a million unread Greek papyri. The new priest who finishes the sentence is a language model named Apollo.
The Austrian Academy of Sciences, with Mistral and Sail Reply, is releasing Apollo, billed as the first advanced LLM for Ancient Greek, trained on about 600 million historical words. It will be free to academics as a chatbot, meant to search and restore tattered papyri and inscriptions—including a worldwide backlog of roughly one million unread Greek papyri. The model switches dialect on sight: Homeric when it sees Homer, Doric when it sees Doric. render_inline_citationcitation_id69
What Happened
Sources are attached to the daily scan.
2026-09-22
San Diego wants AI to draft the reports that send people to prison
This is the memo that decides if you go to prison. California is asking software to write the first draft.
San Diego County is bidding out an AI system to help probation officers write hundreds of pre-sentencing reports a month by summarizing police files, transcripts, and case notes. Those reports are the document judges use when they choose prison, local time, or supervision. The county says the model will not itself recommend a sentence and use will be optional; supervisors are already asking for oversight. Vendor proposals are due October 16. render_inline_citationcitation_id117
What Happened
Sources are attached to the daily scan.
2026-09-22
Robin Williams’ daughter told fans to stop making her dead father a conspiracy puppet
They made Robin Williams argue a conspiracy. His daughter’s review: the voice is robotic, and you’re the puppeteer.
Zelda Williams called out a circulating “private video” of her late father talking conspiracies as obvious, bad AI. She told the human who prompted it to leave him out of “delusional bullshit,” asked who is actually manipulating the public, and closed: “Just because he’s gone does not mean he’s now your puppet. Have some shame.” The family had already revived his Instagram to fight likeness abuse. render_inline_citationcitation_id78
What Happened
Sources are attached to the daily scan.
2026-09-22
DeepSeek is briefing the UN Security Council this week
The UN Security Council put DeepSeek on the agenda. The founder isn’t even going.
Reuters, citing two sources, says DeepSeek will be among firms briefing the UN Security Council on AI and international security this week, with Moonshot also invited. OpenAI CEO Sam Altman plans to brief the Wednesday session; diplomats also expect senior Anthropic representation. DeepSeek founder Liang Wenfeng is not currently expected to attend, and plans could still change. France holds the Council presidency this month. render_inline_citationcitation_id96
What Happened
Sources are attached to the daily scan.
2026-09-22
OpenAI won the Clay version of Navier–Stokes. Mathematicians say that’s the wrong problem.
AI solved a million-dollar math problem. Then mathematicians pointed out it solved the version with a fake wind machine.
Two weeks after 10,000 agents produced a Lean-checked blow-up for Clay’s forced Navier–Stokes option, a new fight landed: mathematicians argue the forced version is not the problem the field actually cares about, and a Thursday paper claims OpenAI’s method can never extend to the unforced equations. OpenAI yesterday stood up an independent IAS-hosted math advisory group (Gowers, Witten, Hairer, et al.) to review and release results—and explicitly said the group will not advise on pacing internal math progress. The same post claims the internal model has now resolved 100+ open problems, without publishing the list. render_inline_citationcitation_id149
What Happened
Sources are attached to the daily scan.
2026-09-22
They’re not watermarks. They’re spymarks.
If it can phone home who made the picture, it’s not a watermark. It’s a spymark.
A short essay that’s ripping through builder Twitter/HN argues Google-style SynthID “watermarks” should be renamed spymarks: hidden signals that can map media back to identity and provenance without meaningful consent. It treats the rename as the intervention—once you say “spy,” the privacy fight is in the noun. This is a coinage/campaign, not a new Google launch. render_inline_citationcitation_id83
What Happened
Sources are attached to the daily scan.