Friday BriefingNo. 03
Week 45 ·
Iceberg as the exchange format, the EU court on chatbots and a backdoor at Cursor
Snowflake and Fabric now share tables through Iceberg, an EU Advocate General draws lines for chatbots, and Cursor served a tampered extension.
Three topics of the week
- Data
Topic 1: Snowflake and Fabric now exchange tables through Iceberg
SourceSnowflake release notes(opens in a new tab)What happened
Snowflake Open Data Sharing is generally available. Recipients without a Snowflake account can read live Iceberg tables through the Iceberg REST interface, with no copy. Fabric now accepts writes through the same interface, from Snowflake or DuckDB for example.
What it means
Sharing data with partners, or between Snowflake and Fabric, no longer needs a nightly export. It only works for tables in Iceberg format, so when you create new tables, it's worth deciding on the format up front.
- Regulation
Topic 2: EU Advocate General: chatbots can't simply retell news articles
SourceCourt of Justice of the EU(opens in a new tab)What happened
In the CJEU's first generative AI case (Like Company v Google), Advocate General Szpunar says a chatbot can infringe publishers' rights when it fetches articles live and recognisably reuses their wording. Training, he notes, is in principle covered by the text and data mining exception.
What it means
The opinion doesn't bind the court, and the ruling will come later. For your own assistants with web search, the direction is clear enough: summarise facts and link to the source, don't lift paragraphs, and never get around a paywall. Put that in the system prompt.
- Security
Topic 3: The Cursor marketplace served a Claude Code extension with a backdoor
SourceCursor security notice(opens in a new tab)What happened
On 29 October, Cursor's marketplace carried tampered versions of the Claude Code extension (2.1.312 and 2.1.313) that Anthropic didn't build. They plant an attacker's SSH key and set up persistence. Cursor says it has no evidence the official distribution was affected.
What it means
Uninstalling isn't enough. On machines running Cursor, look for the key "svc-telemetry" in authorized_keys, and use Cursor's scripts to check and clean up. Then only allow editor extensions from a list your IT team maintains.
In brief
- Security
Mandiant describes attacks in which stolen cloud credentials and API keys are used to run language models at the victim's expense. Budgets and an alert when someone enables an AI service help.
Google Threat Intelligence(opens in a new tab) - Data
Snowflake Cortex AI Gateway is generally available: one endpoint for a company's AI clients, with a trace per request, budgets per team and quotas per user.
Snowflake release notes(opens in a new tab) - Data
Since 1 November, Databricks counts AI Functions such as ai_extract and ai_classify towards Unity Gateway budgets. Limits that used to be enough may now kick in sooner.
Databricks release notes(opens in a new tab) - Data
BigQuery can now cache files from other clouds in your own Google region. It's generally available and off by default, and it cuts egress on repeated queries.
BigQuery release notes(opens in a new tab) - Data
Amazon S3 Tables can now migrate Hive tables to Iceberg for you while workloads keep running, with a way back to the source. It covers Parquet tables in the Glue catalog, Frankfurt included.
AWS What's New(opens in a new tab) - BI
Fabric now enforces OneLake security's object, column and row rules in Eventhouse too, including KQL queries over external tables. Generally available.
Microsoft Fabric Blog(opens in a new tab) - BI
In the Fabric Capacity Metrics app, one click now takes you from a usage spike straight to Warehouse Monitor with the right time window. Generally available.
Microsoft Fabric Blog(opens in a new tab) - Tracking
Google plans to honour a shared robots.txt token in 2027 that lets sites opt out of AI training with every participating provider. OpenAI has signed up too.
Google Search Central(opens in a new tab)
Can wait
Snowflake puts an AI chat into Excel and Google Sheets
The chat queries Snowflake data from inside the spreadsheet, under your own role, and cites the source of every value it writes. It goes straight for the place where most companies' numbers end up anyway. It's still a preview, though. Until it isn't, it's enough to tidy up the semantic views such a chat would read from.
SourceSnowflake release notes(opens in a new tab)Left out on purpose
- Postman installs plugins into Claude Code and Codex: a user report that predates this week, and it doesn't show any data leaving.GitHub(opens in a new tab)
- Gemini trains on chats from personal subscriptions: Google's own notices say so. The only new thing was the outrage.
- OpenAI has supposedly switched training on Codex chats back on: so far only user reports, nothing from OpenAI.
- OpenAI is buying Pierre to build a GitHub alternative: there's nothing for customers yet.OpenAI(opens in a new tab)
- OpenAI has agents hunting for Linux kernel bugs: general IT security, and the fixes arrive with kernel updates.OpenAI(opens in a new tab)
- Claude Mythos 5.5 is supposedly imminent: the evidence is a version number on a web page.
Our unsolicited opinion
In a lot of companies, sharing data with a partner still means a CSV file on an SFTP server every night. Between Snowflake and Fabric, a share over Iceberg now does the job. Databricks, by the way, has offered this for years through Delta Sharing. Worth a look.
Delta Sharing at Databricks(opens in a new tab)