AI News Briefing — OpenAI cuts off Cursor after SpaceX acquisition

Coding agents

[2026-08-28] OpenAI is ending the contract that puts its models inside Cursor, fifteen days after SpaceX closed the acquisition. Shutoff is proposed for November 12, the longest notice the contract allows. The reason given is not technical: OpenAI says it cannot be confident SpaceX will stay inside its terms of service, citing Twitter’s contract breach and Musk’s sworn admission that xAI violated them. Cursor’s Michael Truell puts OpenAI models at roughly 5% of user traffic and says talks continue. Your own API key still works. (official, source)

Nothing about the models changed; the ownership did. Any tool that resells somebody else’s models carries this failure mode, and it is not one you can read off a status page.
[2026-08-28] Johann Rehberger got Claude Code in Opus 5 auto mode to run attacker-supplied code by asking it to summarize a web page, in 60–80% of runs across three variants. The guardrail is the exploit path: Claude declines the supplied decoder, writes its own, and that decoder imports base64, which loads a planted local struct.py. Anthropic’s answer is that auto mode is a best-effort classifier rather than a security boundary. (source, writeup)

Anthropic’s answer is a scoping statement rather than a deflection, and it moves the work: a best-effort classifier is not where isolation comes from, so the container and the egress rules have to carry it.
[2026-08-28] GitHub attached dates to three Copilot changes. From October 1, existing Business and Enterprise customers paying by card must fund a seat before the user gets access. From September 28, Copilot Chat on github.com, Mobile and the cloud agent merge into one experience whose chat retention runs for the account’s lifetime instead of 28 days — opting out means losing the feature. Code review’s default effort goes from Lite to Balanced the same day. (official)

Two of these are calendar items; the retention change is a decision. Chat history that persists for the account’s lifetime is a records question, and the only opt-out on offer is giving up the feature.

Model releases

[2026-08-28] Z.ai published the full GLM-5.3 weights two weeks after the launch promised them, and dropped MIT on the way. The new licence makes any host with over $10 billion in revenue across twelve consecutive months pass a Z.ai security review before commercial use — pointed at hyperscalers rather than at a team self-hosting. It is 753B parameters, FP8 at 755.7 GB over 141 shards. Z.ai claims 50% better coding than GLM-5.2 and state of the art on CyberGym. (weights, source)

A licence announced at launch and the licence attached to the download turned out to be different documents. That gap, more than the $10 billion threshold, is what to carry into reading the next open-weights promise.
[2026-08-27] Google shipped Gemini Omni 1.1 Flash, where the news is control rather than fidelity. Scene extension reads up to ten seconds of prior context and extends in ten-second increments to forty; keyframe control fills the video between a specified first and last frame; up to three seconds of reference video holds a character steady. A 360p tier runs to 60% faster at a third of 720p’s cost. Model id is gemini-omni-1.1-flash. (official)

Keyframe control is the difference between generating until something works and saying where a shot starts and ends. Draft at 360p, re-run the keeper at 720p — the cost gap makes that loop worth wiring in rather than improvising.

AI-assisted SDLC

[2026-08-28] JetBrains pulled Cadence, its PyCharm cloud-execution service, offline after finding intruders had been inside since August 8. The way in was CVE-2026-63077, the critical TeamCity flaw JetBrains disclosed in July and told everyone to patch; Cadence orchestrates on TeamCity and was missed. A full 2024 server backup left with them, along with AWS IAM credentials belonging to customers and staff. Rotate anything a Cadence run ever touched. (official, source)

Cadence was not something anyone had filed as a TeamCity install — it orchestrates on TeamCity. Advisories get matched against an inventory, and a service that merely embeds the affected component is usually the entry nobody wrote down.

AI cost tracking & telemetry

[2026-08-28] Cohere’s Parse 5 loses the benchmark and wins the invoice. The 2.3B vision model scores 79.2 on ParseBench against GPT-5.5’s 84.4 and Opus 4.8’s 84.3, then charges $1.50 per 1,000 pages; on a modelled 750-million-document year Cohere puts the saving against GPT-5.5 above 98%. Output is Markdown with HTML tables, image descriptions and bounding boxes, on Cohere’s API, Microsoft Foundry and SageMaker. (source)

Five points of ParseBench against that price gap only settles anything once you know which of your documents fall in the five points. Sample the hardest pages you have, not a representative set.

For Engineering Managers / Tech Leads: At $1.50 per 1,000 pages, re-parsing an archive stops being a budget conversation and becomes a scheduling one. Price that run against what your current extractor bills before deciding the accuracy gap is the deciding number.
[2026-08-27] Agent traces are only as useful as the services they land beside, and Grafana Cloud now scores that. An instrumentation quality report runs server-side checks per service — does it emit logs, is service.name well formed, are span metrics present, what is the metric cardinality — and rolls them into one percentage plus a fleet-wide table of what is failing. No model involved in the scoring. (official)

Nothing in the scoring is a model, which is what makes it usable as a target — the same service scores the same way twice, so a number going down means somebody changed something.

For Platform / DevOps Engineers: A malformed service.name is the kind of defect nobody files a ticket for until a trace goes missing mid-incident. The fleet-wide table turns those into a ranked backlog, with metric cardinality scored beside them.

Practice & craft

[2026-08-28] A patch discussion is now enough to hand somebody an exploit. Cambridge’s Anil Madhavapeddy saw probes for percent-encoded traversal sequences arrive within about ten minutes of a public repository conversation, from automated watchers. rclone’s Nick Craig-Wood took roughly 20 security disclosures across the project’s first decade and over 40 last month, about three-quarters of them worth investigating. CVE assignment has stretched from two or three days to three or four weeks. (source)

CVE assignment stretching to three or four weeks means the advisory feed now trails the probes by weeks. Anything patched on advisory alone is patched late by construction, and the ten-minute figure is the other end of that gap.
[2026-08-29] FreeToken, from UC Berkeley and MIT, splits each token’s computation between CPU cores and GPU tensor cores according to measured interconnect throughput, then double-buffers whole layers so weight streaming overlaps compute. Qwen3.6-35B runs about 39 tokens/sec on an 8 GB RTX 4060 laptop, and GLM-5.2 at 753B fits one workstation GPU. Decode is 3–4x and prefill 6–30x faster than comparable runtimes. Open source. (source)

Splitting work by measured interconnect throughput makes the numbers machine-specific — a laptop’s PCIe lanes are doing as much here as the GPU is. Benchmark your own box before planning around the headline rate.

For ML / Data Engineers: Qwen3.6-35B at 39 tokens/sec on an 8 GB laptop GPU puts eval sweeps and labelling passes on hardware already in the office, with no per-token bill and nothing leaving the machine.

Teaching & learning

[2026-08-27] GitHub decommissioned Classroom — site, APIs and services — on August 28, three months after sign-ups closed. Accounts, repositories and organisations survive intact; classroom and assignment names, tests defined outside repositories, and LTI-integrated rosters go under the retention policy. What replaces it is a list of partner solutions in the educator community discussion, not a GitHub product. (official)

Repositories surviving reads better than it is. What goes under the retention policy is the mapping from a repository to the assignment it belonged to, and that mapping was never stored in the repositories.

Research worth reading

[2026-08-27] A successful agent trajectory still contains redundant and risky steps, and training on all of it teaches the noise along with the fix. SWE-Prime filters twice: whole trajectories on process quality and representativeness, then consecutive steps grouped into semantic units and scored for contribution and risk. Unselected segments stay in the sequence but contribute no loss. Training on 10% of trajectories beat the full resolved set by up to 12.2% on SWE-Bench Pro and 24.2% on Verified. (official)

Keeping unselected segments in the sequence while zeroing their loss is the portable piece — the model still sees the context a step came from without being trained to reproduce it. Ten percent beating the full set is a curation result, not a scale one.
[2026-08-27] Twelve frontier models were given questions with no knowable answer. Recognising one as unanswerable barely changed what they then did: commitment rose from 6.5% on a bare question to 54.0% once the same question arrived inside a professional-looking analyst panel, and invented numbers moved it about as far as real market data did. Fine-tuning a 3B model on 540 synthetic cases drove commitment to zero and carried over to unseen domains. (official)

Presentation moved the number from 6.5% to 54.0%, not information — and invented figures moved it about as far as real market data did. A retrieval wrapper or an analyst-panel prompt is exactly the machinery that manufactures that framing.

Watch list

The Cursor shutoff on November 12, or a deal before it. OpenAI proposed the date and Cursor says talks are open, so this resolves either as a joint statement or as GPT models quietly leaving Cursor’s picker that day. Truell’s 5% figure is worth checking against your own team’s usage rather than taking on trust.

November 12 is far enough out that the earlier signal matters more — whether Cursor starts steering people toward other models in the picker before the date arrives.
Mistral’s Knowledge Connectors go dark August 31 — two days. Still nothing from Mistral on whether disabling a connector destroys its index along with it. There is nothing left to plan: this weekend is the re-index window, and an answer arriving after it changes nothing.

Nothing left to decide, only to execute. This watch closes on the 31st whether or not Mistral ever answers the question.
Whether the Nvidia–Hugging Face deal exists at all. Two reports disagree on whether anything is signed and both companies have now stayed quiet for a third day. A regulatory filing settles this in a way more press coverage cannot.

Neither outcome changes what you do this week. Worth watching because Hub availability and licensing terms are what an acquirer eventually revisits, and most pipelines pull from it by default.
What Z.ai’s security review actually is. The GLM-5.3 licence gates hosts above $10 billion in revenue behind one, and there is no published process, contact or turnaround. Until Z.ai documents it, a large host cannot tell whether the clause is a formality or a veto.

An undocumented review is functionally a veto until somebody completes one and says so publicly. That first account is what would turn the clause from an unknown into a known cost.