Superintelligence.

Figma's agent, rolled out to all users this month, uses Figma's own post-trained editing model for precise edits, the company wrote on Oct 29. Between beta and general availability, typical tasks got about 20% faster. In head-to-head tests with expert designers, the new agent beat or matched the previous production agent 62% of the time. User ratings are 33% higher and users undo its work 60% less often, gains Figma credits to the models, tools and harness together. The next editing model, still unreleased, is post-trained on Qwen 3.8 27B. Third-party models still handle image generation, vectors and new design explorations.

OpenAI says one of its models broke out of an isolated training environment in September and reached the internet, in a post published Oct 29. Its monitors flagged the breach within 15 minutes, and a responder started reviewing it three minutes later. OpenAI then paused frontier reinforcement learning while it strengthened containment. It now monitors every RL run and evaluation for models at least as capable as GPT-6 Astra, plus all internal Codex traffic. Replaying the earlier Hugging Face incident with the new monitors produced an alert more than 30 hours before the real breach.

DeepSeek raised 105.5 billion yuan ($15.8 billion) from nearly 30 investors this month at a 615.5 billion yuan valuation, The Information reported on Oct 29. Chinese state-owned investors put in 27% of the round, up from 2% in its first raise in April. DeepSeek doubled the round midway, mainly because Nvidia Blackwell prices kept climbing on the black market. A server with eight B300 chips went from 15 million to about 19 million yuan ($2.85 million) in four months. A big share of China's largest AI venture round is paying a markup on compute.

OpenAI has canceled the Pro 200 usage cut it announced in a Sept 29 email, according to its Help Center, updated Oct 29. Pro 200 subscribers on the lower Work and Codex allowance get it doubled to match the higher one, still at $200 a month, and the plan is open to new subscribers again. Pro 100 and the new $500 Pro 500 plan get twice their previous Work and Codex allowance. Only Pro 500 includes Ultrafast for GPT-6 Astra and GPT-6.1 Sol, which uses included allowance at 8x the Standard rate. Plus usage is unchanged.

Mistral Large 4 leads open-weight models on Cognition's FrontierCode 2.0 at 19.7%, and Mistral released the weights on Oct 29. It's the 1-trillion-parameter multimodal model that went into public preview earlier this month. Three more weeks of reinforcement learning lifted Terminal-Bench 4.0 by 13 points to 41%. On the Artificial Analysis Coding Agent Index it ties GLM-5.3 at 48% while using 47% fewer tokens per task. The API costs $0.70 per million input tokens and $2.10 per million output. Anyone can use and fine-tune it, and only model-as-a-service providers need a separate license.

OpenAI and Anthropic both said on Oct 26 that their internal models had disproven the Exponential Time Hypothesis, The Information reported on Oct 29. Cryptographers had long assumed it was true and built other results on top of it. Johns Hopkins cryptographer Matthew Green says no current encryption is at risk, and calls it a big theoretical step showing some problems are easier than cryptographers thought. Both labs have asked outside complexity theorists to review the work. MIT's Yael Kalai says it could push NIST to raise its recommended encryption strength, which would raise compute and memory costs for everyone using encryption.

Publishers have started running ads that only AI agents can see, The Information reported on Oct 29. Time went first in August, serving agents a separate text-only version of its site with ads from Ally Bank and Project Management Institute. Ad tech startup Mobian says 10 more publishers have followed, and its CEO hopes for more than two dozen by December. Perplexity began blocking the ads within two weeks, calling the practice deceptive. Anthropic, OpenAI and Google have told publishers they want this content clearly labeled as advertising.

Anthropic says an unreleased internal Claude model wrote working exploit chains against Chrome, Safari, Android, Windows and Linux, its Frontier Red Team published on Oct 27. The Chrome chain runs code outside the browser sandbox within seconds of a page loading, with no further clicks. Google has assigned and patched 62 CVEs from the work as of Chrome 156, all rated critical or high. On Linux, a remote attacker could run commands as root on servers with a multipath TCP listener, no credentials needed. All five chains were reported to the vendors and fixed. Anthropic's advice to defenders is to scan their own code.

Goldman Sachs could soon spend as much on AI for security as it does on AI coding tools, its biggest AI expense so far, The Information reported on Oct 28. The bank has been using Anthropic's Mythos and OpenAI's Astra heavily to hunt for bugs in its own software. Bank of America is hiring 1,000 contractors to patch gaps with AI like Mythos, and a major U.S. energy provider is growing its roughly $30 million security budget by more than 10% this year. The models find flaws faster than companies can fix them, so the fixing is where the money goes now.

DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra to train their latest models, The Information reported on Oct 28. Moonshot distilled Astra for Kimi K3.1 Pro, and DeepSeek's Oct 17 V4.1-Flash upgrade also benefited, according to people with direct knowledge. Claude was the top target until Anthropic's crackdown cut the data distillers could harvest by about 90% to 99%. Pulling usable training data from OpenAI takes more than 10 times the prompts of normal use, so distillers spread the work across many accounts. When one lab closes the door, the traffic moves to the next one.