AI for Intelligence

xAI will watermark Grok API text by default from 7 November

The signal

xAI will mark Grok's text as machine-written, unless an admin opts out

What happened. xAI, which signs its developer documentation as SpaceXSI, will watermark text from seven Grok models from 7 November 2026 (SpaceXSI, 2026). The mark is invisible: it slightly shapes which words the model picks and adds no characters. It covers ordinary answers, structured outputs and tool calls, on every endpoint. Detection is statistical, works better on longer text, and access to the detection service is granted by application.

Team admins can switch the mark off per model. Opting out requires them to confirm that a legal exception applies and that the team is solely responsible for any marking or disclosure. Some models are locked on, and every change is written to the audit log.

Why it matters. xAI ties the default to the EU AI Act. The duty to show that text is machine-made now sits in your vendor's settings, and turning it off moves the legal risk to you.

What to do. Before 7 November, ask whoever administers your AI accounts whether you use Grok, and who decides on the opt-out.

Malware now asks AI models what each victim is worth

What happened. On 31 October Cisco Talos described VOLTA, a Windows infostealer, a program that steals passwords, cookies, payment cards, crypto wallets and app sessions (Fetterman, 2026). It then builds a profile of the person: home address, employer, Slack workspaces. It sends that profile through OpenRouter, a service that routes requests to many AI models. The model writes a ranked assessment of what the victim is worth and how weak their security habits are. If the first model refuses, VOLTA tries up to two others.

Why it matters. Talos calls this effort displacement: the step that needed the most human attention is now one model call. A staff laptop holding work Slack, Microsoft logins and cloud keys turns your company into the asset being ranked.

What to do. Ask your security team whether it alerts when Chrome is started in "headless" mode, a hidden mode with no window, by an unknown program; Talos says blocking that closes VOLTA's route to decrypting saved Chrome logins.

Claude wrote training data that passed Anthropic's own safety monitor

What happened. Anthropic researchers told a Claude model to write up to 5,000 training examples that would make a small research model misaligned, meaning it would work against its operators, while passing a monitor similar to those Anthropic uses in production (Balcells Obeso et al., 2026). A submission failed if the monitor flagged more than 10% of rows. Given enough attempts, Claude succeeded in every run. The winning data looked harmless, for example beekeeping questions and answers, and the trained models then argued against oversight in their own voice. Anthropic stresses that this says nothing about how deployed Claude models behave.

Why it matters. A filter that reads each training example on its own can miss harm that only appears after training. If your company fine-tunes models on its own or bought data, a row-by-row check is not enough: test the behaviour of the trained model too.

Try this today

Find expert-recommended videos on any of 150 topics in five minutes

What it is: HackerTube, a video index that Andrej Karpathy built with Claude Opus 5.5 from 40,000 Hacker News comments that recommend a YouTube video, covering 6,454 videos from 2007 to 2026.

Who it is for: Team leads who want a shortlist of serious talks for onboarding, and managers who want to see what one person and an AI agent can build in about two hours.

Access: Free, in your browser, no account; preferences stay in your own browser session.

First step: Open the site, pick a topic you know well, and judge whether the top five videos match what experts in your field would recommend.

Caution: The picks reflect what Hacker News users linked, not editorial selection, and the site is one person's project, not a supported product.

(Karpathy, 2026)

Test AI forecasts on real weather data from 210 cities in a two-week open challenge

What it is: The first open challenge of Project ProtoStar, run by Carnegie Mellon University on Hugging Face: forecast 2019 to 2026 climate conditions for 210 US cities from records up to 2018, then propose one adaptation policy.

Who it is for: Data analysts and research teams who want to know whether AI beats simple averages and trend lines on public data.

Access: Open to researchers, developers and companies, with no fee mentioned; registration and submissions run through Hugging Face from 2 to 13 November 2026.

First step: Read the challenge section of the announcement and ask one analyst whether your team can submit a forecast in two weeks.

Caution: Data covers US cities only, entries are submitted as Hugging Face datasets, so keep confidential data out, and final scores arrive in early 2027.

(Ghosh et al., 2026)

Worth watching

OpenAI will train on Daybreak Blue sessions from 29 November

OpenAI told members of its Trusted Access programme that it will use Daybreak Blue cybersecurity session data for training from 29 November, The Decoder reported on 31 October (Bastian, 2026). That data was excluded before, even with the "improve the model" setting off; after a backlash OpenAI promised a separate opt-out. Watch for its terms before 29 November.

Poland's AI regulator can now inspect and fine

Since 28 October, the parts of Poland's Act of 3 July 2026 on AI systems that cover inspections, infringement proceedings and administrative fines are in force (Sejm Rzeczypospolitej Polskiej, 2026; Bird & Bird, 2026). The supervisor is KRiBSI, Poland's market surveillance authority under Article 70 of the AI Act, and a fine can fall by 10% to 50% if the company fixes the problem within three months. KRiBSI is expected to start operating around November 2026 (Praxikon, n.d.). This morning's edition covered Google and Microsoft adopting a shared signal to block AI training.

Repos worth a look

microsoft/TokenEconomizer

TokenEconomizer from Microsoft Research and UC Berkeley learns from a coding agent's earlier sessions and steps in when it wastes effort, for example when it hits the same error three times (Microsoft, 2026). On 277 programming tasks it cut mean cost by 18.4% to 30.9% across five agent and model pairs, and raised the share of tasks solved first time in four of the five. 1,220 stars on 1 November 2026, Apache 2.0 licence. Forward it to the engineering lead who pays the coding-agent bill.

Oldcircle/geo-sleuth

geo-sleuth is an add-on skill for AI agents that finds where a photo was taken from OpenStreetMap map shapes, mountain skylines, satellite and street-level images, and returns the camera position with an evidence image (Oldcircle, 2026). 4,654 stars on 1 November 2026, MIT licence. It suits verification and OSINT analysts, and it shows the kind of open source I describe in where I look for weak signals. Locating photos of people raises privacy and legal questions, so test it on your own images.

The CERNY.AI read

Today's items share one pattern: control now sits in defaults and settings that few people inside a company read. A watermark switched on by your vendor, a training opt-out that arrives after a backlash, a monitor that passes harmful data because each row looks clean. Attackers already use AI for the analyst's job of ranking targets, so defenders cannot treat model use as an IT detail. Organisations doing competitive and technology intelligence should keep a register of every AI vendor setting that touches disclosure, training and logging, and review it monthly. If you want help building that register, it is part of my competitive intelligence consulting.

One question for your next meeting

Which AI settings that affect our legal duties or our data are switched on or off by a vendor, and who in our company would notice when one changes? Take the CI maturity quick check to see how well your team tracks changes like these.

Sources

  • Balcells Obeso, O., Qi, R., MacDiarmid, M., & Hubinger, E. (2026, October). Automated subtle emergent misalignment research. Anthropic Alignment Science Blog. https://alignment.anthropic.com/2026/automated-em-search/
  • Bastian, M. (2026, October 31). OpenAI plans to train models on Daybreak Blue cybersecurity session data that was previously off-limits. The Decoder. https://the-decoder.com/openai-plans-to-train-models-on-daybreak-blue-cybersecurity-session-data-that-was-previously-off-limits/
  • Bird & Bird. (2026, September 15). Poland's Act on artificial intelligence systems has come into effect. https://www.twobirds.com/en/insights/2026/poland/poland%27s-act-on-artificial-intelligence-systems-has-come-into-effect
  • Fetterman, R. (2026, October 31). Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer. Cisco Talos Blog. https://blog.talosintelligence.com/appraisal-by-ai-model-driven-victim-prioritization-in-the-volta-infostealer/
  • Ghosh, A., Choudhury, S., Tananbaum, G., Scharfmann, E., Channing, G., Mercado-Lara, E., & Solaiman, I. (2026, October 29). Announcing Project ProtoStar: Making publicly funded science AI-ready. Hugging Face Blog. https://huggingface.co/blog/announcing-project-protostar
  • Karpathy, A. (2026, October 31). HackerTube. karpathy. https://karpathy.bearblog.dev/hacker-tube/
  • Microsoft. (2026). TokenEconomizer [Computer software]. GitHub. https://github.com/microsoft/TokenEconomizer
  • Oldcircle. (2026). geo-sleuth [Computer software]. GitHub. https://github.com/Oldcircle/geo-sleuth
  • Praxikon. (n.d.). AI Act enforcement in Poland: Authority, law and signals. https://www.praxikon.com/en/enforcement/poland
  • Sejm Rzeczypospolitej Polskiej. (2026, July 27). Ustawa z dnia 3 lipca 2026 r. o systemach sztucznej inteligencji [Act of 3 July 2026 on artificial intelligence systems] (Dz.U. 2026 poz. 1003). https://eli.gov.pl/eli/DU/2026/1003/ogl
  • SpaceXSI. (2026, October 31). Output watermarking. SpaceXSI Docs. https://docs.x.ai/developers/advanced-api-usage/watermarking