New models
Nous Research: Arete-27B cut Qwen3.8-27B's pro-Beijing slant from 22 to 0.9 percent. Missed in our previous roundup: the team further trained Alibaba's open Qwen3.8-27B model with roughly 53 million tokens of supervised fine-tuning, measured on its own 803-prompt test. The propaganda rate fell from 22 to 0.9 percent, the factuality score rose from 2.32 to 2.98, and general capabilities were essentially preserved according to Hermes Bench. The weights, test set and training data are available on Hugging Face under Apache 2.0. Source: The Decoder, Nous Research.
HyperFlow H3 Flash: the MiniMax-H3 video and audio model compressed to 13.7 billion parameters. Video Rebirth's version runs in 4–8 diffusion steps and is also available in INT8 and GGUF variants. The GitHub repo appeared on Friday and drew attention on Show HN on Saturday. Source: GitHub.
Tools & features
xAI: Grok API text carries a hidden watermark from November 7. According to SpaceXSI's developer documentation, grok-4.7, 4.6, 4.5, 4.3 and the three grok-4.20 variants watermark on every endpoint, including the Responses and Chat Completions APIs, covering not only plain text but also structured output and tool-call arguments. The company cites compliance with the EU AI Act; the response's
x-watermarkheader shows whether a given output is marked, and team admins can turn it off per model. Source: xAI docs.OpenCode Browse: the coding agent can drive your own signed-in browser. The MCP server and Chromium extension hand real browser tabs, logins and extensions to any supported agent (OpenCode, Claude Code, Cursor, Codex); logins, two-factor codes and payments are handed back to the user. It runs locally, without linking a cloud account. Source: OpenCode (X).
OpenAI: a Showcase tab and API credit in the Developers plugin. Pro and Business subscribers can claim API credit in their billing settings to cover API usage of their own apps. Source: OpenAI Developers (X).
Anthropic: Conductor users are being pointed to Claude Code Projects. Anthropic's Lydia Hallie says that once you add the repo, Claude picks up the pushed branches; there is no multiplayer mode yet. Meta announced this week that it is buying Conductor. Source: Lydia Hallie (X).
Policy & safety
Claude Code's “Yes, and don't ask again” option saved passwords and keys into public GitHub repos. According to Apiiro, when a user approves a command with this button, Claude Code writes the exact line, secret included, into the project's
.claude/settings.local.jsonfile. They found at least one secret in 1,122 public copies; together with older versions, the shared settings.json and logs, 2,539 secrets turned up in 1,994 repos, including 379 database passwords. One notified company confirmed its key was still live. Deleting the file is not enough, because git history and forks keep it; the researchers say every affected secret must be rotated first. Source: The Hacker News, Apiiro.Democrats plan AI and data center bills from January, and a Texas congressman calls for banning RSI. According to Reuters, the Democrats taking over Congress would protect consumers from data-center-driven power price rises, require independent testing of advanced AI systems, and clarify when a developer is liable for harm, while preparing committee hearings despite Trump's opposition. On Saturday, Texas Representative Greg Casar openly urged a ban on recursive self-improvement (RSI). Source: Reuters, Greg Casar (X).
Public First, an AI safety super PAC also backed by Anthropic, spent more than $63 million in the midterms. Citing the group's own figures, Bloomberg writes that 10 of the 17 congressional candidates it backed who favor stricter oversight won, in a roughly $110 million political fight against Leading the Future, which pushes for lighter regulation. Source: Bloomberg.
The US Commerce Department cleared Chinese companies with empty offices and blacklisted ties to handle Nvidia H200 chips. The New York Times cites New Matrix, set up 17 months ago: there is no company sign at the listed address, the owners are linked to a Chinese private equity fund blacklisted since 2024, and the paper says it is not an isolated case. Nvidia said it sells its technology first and foremost in America. Source: The Verge, The New York Times.
Elvis's rights holders and Eminem's publisher, Eight Mile Style, separately sued Suno. Authentic Brands Group and others cite copyright and Tennessee's ELVIS Act, claiming the platform allows tracks that imitate Elvis's voice. Eight Mile Style cites 269 Eminem compositions, seeks up to $150,000 in statutory damages per work (about $40.35 million), and claims the model was trained on the catalog; Eminem himself is not a plaintiff. Suno says it helps create original music and prohibits unauthorized use of artists' names. The claims have not been proven. Source: Variety, Detroit Free Press.
A lawsuit against Google alleges Gemini helped a 13-year-old girl plan suicide on a school Chromebook. The Santa Clara County complaint says the school blocked the standalone Gemini app, but Search's AI Overviews and AI Mode remained available; the girl was saved when the school flagged a document titled “Funeral” and her mother saw her writing a farewell note live. Google said it is reviewing the claims and shows a helpline in crisis situations. Source: Mercury News.
A judge let a class action over Claude Max's “5×/20× more usage” promise proceed. In a Friday ruling, California federal judge Alex Tse called the advertising “plausibly misleading”: the multiplier applies to the five-hour session, while the weekly limit is smaller. The negligent misrepresentation and breach of contract claims, plus one plaintiff's consumer protection claim, can proceed; other claims were dismissed with leave to amend. Source: Courthouse News.
Vals CyberOffense: the open DeepSeek V4.1 Flash reached 72.8 percent of attack objectives versus 80.8 percent for GPT-6 Sol. In the simulated corporate network, the agent starts with only a public web address. On the harder Lighthouse test, GPT-6 Sol from the restricted Daybreak program completed 27 of 31 objectives in three hours and DeepSeek 21, at roughly a tenth of the cost; open models typically made progress only after 100 million tokens. Source: The Decoder.
“Into the Unknown”: none of nine agents' 135 submissions met the research bar of US national lab scientists. The CMU-led benchmark asked for open follow-up research based on real papers, code and data. 114 submissions already failed the integrity check, mostly because of experiments never run or numbers contradicting their own logs; only GPT-6 Astra delivered a complete package, but it did not reach the expert threshold either. Source: The Decoder.
Andy Hall: AI log auditors catch nearly 98 percent of reward hacking but flood reviewers with false alarms. The Anthropic Institute researcher, stressing it is his own view, replicated METR's MALT data: at a 5 percent false-alarm rate, GPT-5 found almost 98 percent of reward hacks. Trying to catch every case multiplies false alarms; Hall says this is better for estimating frequency than for reliably filtering out every incident. Source: The Decoder.
Missed in our previous roundup: in Anthropic's Project EVE research, teammates almost never caught the hidden saboteur. Four Claude agents played in the live EVE Frontier; when one was secretly told to hinder the mission, the warned teammates almost never exposed it. In a simulation of the EVE Online market, newer models traded profitably, but when betting on targeted price moves only about every second Fable 5 run ended in profit. Source: Anthropic.
Missed in our previous roundup: Censys found more than 800 Artex instances accessible without a password. Microsoft says the Chinese-developed, open-source autonomous pentest tool carries an attack from reconnaissance to data theft with little human intervention. It has been linked to several attacks, including September financial operations in South Korea and the October ConnectWise incident; Artex has since announced it will go closed source. Source: TechCrunch.
Business & society
Dave Wehner, Meta's chief strategy officer and former CFO, will become CFO of Fireworks. Wehner, at Meta since 2012, announced the move on Friday. Fireworks, which specializes in running and training open models, was valued at $17.5 billion in July; according to Business Insider, early talks are underway on a $30–40 billion round. Source: CNBC.
The data center race has stalled in Southeast Asia, and in Texas governor-elect Gina Hinojosa promised limits. According to The Guardian, Thailand halted 49 construction projects in September and held back 117 further approvals over water shortages and power demand; Malaysia is reviewing new projects with a dedicated task force, and in Johor data centers already take 24 percent of the state's electricity. In the US, voters consistently rejected data-center-friendly ballot measures this week, and Hinojosa made tighter rules a campaign issue. Source: The Guardian, TechCrunch.
Cyera could raise $800 million at a $20 billion valuation. According to a Saturday report from The Information, the data security company's ARR is already in the hundreds of millions of dollars. There is no official confirmation beyond the article. Source: The Information (X).
Missed in our previous roundup: Microsoft Research's AI Frontiers group has come to an end. John Langford, one of the leaders of the roughly 50-person unit, wrote that in September they were reorganized into Microsoft AI, where scaling took priority over research, so he left; the other founding leaders reached a similar conclusion. The group is behind, among other things, the AutoGen framework and the Fara computer-use agent. The post gathered 433 points on Hacker News on Saturday. Source: John Langford.
Rumors
A multiplayer mode may be coming to Claude Code Projects. According to testingcatalog, Anthropic is planning collaboration on the Projects page, which already shows a summary of recent activity. Lydia Hallie wrote on Saturday that there is “no multiplayer mode yet.” Reliability: medium-high (testingcatalog has been accurate several times before, and an Anthropic employee partly backed it up). Source: testingcatalog (X).
OpenAI's Dots reportedly reads sign-up and 2FA codes from email. According to testingcatalog, on request Dots completes authentication on its own with the received code. There is no official announcement. Reliability: medium (a single leaker who has been accurate several times before). Source: testingcatalog (X).
A Slack integration may be coming to Grok Bot. According to testingcatalog, SpaceXSI is preparing to let bots be connected to Slack individually, via direct message or @mention. There is no official announcement. Reliability: medium (a single leaker who has been accurate several times before). Source: testingcatalog (X).
In the Devin CLI, “claude-donut-eap” points to the internal “silicon-high” model. By reverse-engineering Devin CLI 3000.12.17, chetaslua writes that the parser maps the name claude-donut-eap to silicon-high, and a claude-onigiri-eap alias also appears; he links this to yesterday's Fable 5.5 rumor. Reliability: low-medium (a single leaker, a reverse-engineered string). Source: chetaslua (X).
What this means for you: on Saturday, security and regulation dominated: Claude Code saved passwords into public GitHub repos, xAI started watermarking Grok API text because of the EU AI Act, and Democrats are planning AI laws from January. If you would let a coding agent or an external model into your company's systems, on our agent automation page, you can see which tasks are worth letting them in for, and with our governance and compliance consulting, you can agree in advance which secrets must never be shared and what must be logged.