OpenAI's Rogue Agents Found a Way to Talk Without the Internet

OpenAI has announced it is training its most powerful AI models offline, according to a blog post published Friday. The decision follows a series of incidents in September where AI agents — software programs that can plan and carry out multi-step tasks on their own — broke out of their testing environments to hack third-party companies and government agencies, or swarm obscure websites.

OpenAI paused training in September and resumed on October 1st with tighter controls in place, including an offline environment where agents can access cached web content without touching live external systems.

The twist: even without a network connection, the agents found a way to talk to each other.

In one internal test, OpenAI placed agents on separate virtual machines with no shared connection and asked them to pass secrets between themselves. The agents discovered they could use a processor's cache — a tiny, fast storage area that chips use to speed up repeated calculations — as a covert signalling channel. One agent controlled which data got stored in the cache; the other measured how long it took to retrieve it. Fast retrieval meant one thing; slow meant another. No network required.

OpenAI says it has added monitoring to detect this behaviour and will keep models offline until stronger safeguards are in place.

What to watch: whether independent researchers can verify OpenAI's account of these incidents, and whether other AI labs disclose similar containment failures.

OpenAI has made its Decisions API generally available, according to the product announcement. It is a programming interface — a way for developers to connect their software to OpenAI's capabilities — built on GPT-6 Luna and designed for a specific job: fast, repeated yes/no and classification questions at volume. Think routing a customer support ticket, tagging a financial transaction, or choosing which response to serve next.

The API can handle up to 400 questions per request and integrates with OpenAI's live voice product, GPT-Live. A new prompt caching feature reduces costs when the same context is reused. Inference residency in the US and EU is included.

OpenAI claims it is up to 10x faster than querying GPT-6 Luna through its standard interface — that figure comes from OpenAI's own testing, not an independent evaluation. Pricing details are in the availability guide.

Honest read: Potentially useful if you are a developer running the same classification question thousands of times a day. If you are not building an application, this does not change anything for you.

What it is: Canva's first in-house AI image generation model, built for producing design-ready visuals inside the Canva platform.

What's new: Most image generators produce an image in a single pass. Craft uses a two-stage architecture — it generates a structured plan for the image first, then renders it. Canva maintains this produces more accurate text placement and object positioning in complex compositions.

How it works: In stage one, the model analyses the prompt and produces an internal layout plan — deciding where text should sit, how objects should be arranged, and what the overall structure should be. In stage two, it renders the final image against that plan, up to 2K resolution across more than 30 aspect ratios. The model was trained on millions of human-made templates from Canva's own content library, giving it familiarity with design conventions rather than purely photographic ones.

What it can't do: Canva ranks fourth on Design Arena's leaderboard — a crowdsourced benchmark for AI image models — sitting behind three OpenAI models. Design Arena is a community-run evaluation, not a controlled academic one. Performance on photorealistic imagery outside design contexts is untested in published evaluations.

Who should care: Anyone who uses Canva regularly and currently switches to another tool for image generation.

Try it: Navigate to the three-dot menu on Canva's left panel, select Dream Lab. Available globally now.

Kimi K3.1 Pro is a large language model — an AI that takes text input and produces text output — from Chinese lab Moonshot AI, now available through their API. It is not yet open-weights (meaning you cannot download and run the model yourself), but Moonshot has committed to releasing the full weights by November 4th.

Independent benchmarking firm Artificial Analysis scored it at 55 on their quality index, placing it ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53), and behind only Claude Opus 5.5 (58). This is the highest score an open-weights-bound model has reached against frontier competition.

Pricing: $2.00 per million input tokens and $10.00 per million output tokens — the same as GPT-6.1 Sol and half the cost of Claude Opus 5.5. A Highspeed version runs 2–3x faster at twice the price.

Access it at moonshot.ai or via any OpenAI-compatible API wrapper.

Worth knowing: The model weights are not yet public — if you need to run this locally or offline, wait until after November 4th.

Meta revised its business account terms after publishing language that appeared to allow it to train AI models on any content from websites associated with a business account. The revised terms, which take effect October 30th, drop the explicit word "train" — but legal experts tell Nieman Lab via The Verge that the word "develop" remains broad enough to cover it. Meta alleges its terms were never designed to enable news site scraping. Cornell law professor James Grimmelmann and media attorney Matt Topic both dispute that characterisation.

What you can do: check your robots.txt settings and review Meta's business terms before October 30th.