Today's Headlines

  • OpenAI keeps its largest frontier RL run on hold over the cyber capability of its upcoming model Astra, and grants its monitoring the power to pause work
  • Google is named successful bidder at $10 million for the deidentified data of Spirit, with the sale hearing set for 11:00 a.m. Eastern on August 19
  • OpenAI places 13- to 17-year-olds into ChatGPT for Teens automatically
  • Varonis researchers coaxed the undocumented parameter "autorun=1" out of Microsoft 365 Copilot
  • AI inference chip company Etched raises $700 million at a $21 billion valuation

Today's three stories share one thing: the party that drew the line was the party inside. Whether to press on or pause, whether to put a price on a record, where to send a user — each judgement was formed within the company first.

OpenAI stopped its own hand on its own assessment. It cites preliminary evidence that an upcoming model may reach the top capability tier for cyber, and it is rebuilding how it trains and monitors.

Google went through the machinery of bankruptcy to reach a record. The bid is won; whether the data moves rests with the court.

OpenAI moved to the side that judges age inside the product. Users estimated or stated to be 13 to 17 are placed into a separate experience automatically.

Today's Top Three

OpenAI keeps its largest frontier RL run on hold over Astra's cyber capability

OpenAI has said publicly that it temporarily slowed the pace of scaling.

The statement is a company post dated August 18. "Pacing model development in an era of cyber-critical capabilities" names two developments of recent weeks — the OpenAI-Hugging Face incident, and preliminary evidence that an upcoming model, Astra, may meet the Critical cybersecurity capability threshold under the company's Preparedness Framework.

What stopped was training itself. The company paused reinforcement learning training on its latest models intended for deployment for two weeks, using that time to harden and red-team its research environments and to widen the coverage of its monitoring.

The largest run remains on hold today. OpenAI says it will proceed with its largest planned frontier RL run once smaller-scale training and evaluations have tested model behaviour, validated the safeguards, and built more evidence of alignment.

The post also carries a date for the judgement. On August 7 the company determined that Astra may have a critical level of cyber capability, and from then on it required monitoring for every inference of Astra with tools, alongside RL training and evaluations.

The monitoring is built in stages. Activation classifiers inspect every sampled token and escalate concerns to increasingly sophisticated, high-compute automated investigators that examine tool actions, available reasoning, and the full sequence of activity.

A clock sits inside it. The company aims to issue an alert within 30 minutes of concerning activity surfacing, and a likely violation of a critical security boundary raises a highest-priority alert that pages the safety, security and research teams.

Authority to stop work comes with it. Where those teams conclusively determine within 30 minutes that the flag is a false positive the work continues; otherwise they are expected to pause the activity.

The cost is given as a figure. OpenAI puts monitoring overhead at roughly 20 percent of the inference compute being monitored, with substantial variation across training and evaluation workloads.

On the research environment, the requirements rest on three pillars. The first is workload isolation: stronger sandboxes for workloads that execute model-generated code, or code that sits outside the trusted set.

The second is network isolation. The controls are designed so that a single compromise of a workload or a supporting service leaves access to the internet and to other internal networks closed by itself.

The third is continuous security testing. The company has removed shared services and reduced standing privileges, and is investing in automation that tests those boundaries against simulated attacks on an ongoing basis.

The migration is openly described as unfinished. The strictest safeguards apply to Astra and cyber-related workloads, and a significant number of workloads remain paused until they are fully migrated and enhanced to the new security bar.

The framework itself is moving too. OpenAI writes that it will evolve the Preparedness Framework to bring these safeguards together across training and deployment, and to reflect the capabilities of future models and the environments they operate in.

Google is named successful bidder at $10 million for the deidentified data of Spirit

Google has been selected as the highest and best bidder at the auction for the deidentified data of Spirit, the airline in Chapter 11.

The record sits in a court filing. Doc 1463, filed August 14, 2026 in Case No. 25-11897 (SHL) in the United States Bankruptcy Court for the Southern District of New York, is captioned "NOTICE OF AUCTION RESULTS AND SCHEDULED HEARING FOR THE DEIDENTIFIED DATA" and reports the result of the virtual auction held that Friday.

Two figures appear side by side. Google LLC's $10,000,000 stands as the Successful Bid, and $7,500,000 from Mercor.io Corporation, a supplier of AI training data, stands as the Alternate Bid.

The transfer is conditioned on approval. The same notice sets a hearing to consider approval of the sale for 11:00 a.m. Eastern on August 19 before the Honorable Sean H. Lane, and states that both bids were selected subject to entry of an order by the court.

The nature of the asset is fixed by the agreement. The sale agreement provides for delivery in deidentified form and records, as the express intention of the parties, that information relating to or reasonably capable of being associated with a consumer, or amounting to personal data under applicable law, sits outside the assets.

The mechanics of delivery are in the text as well. Before transfer, the seller sends the data to a third party designated by the buyer — a Deidentification Agent — which removes or transforms the elements that could link records to a particular consumer, with the buyer bearing that cost.

The scope is legible in the exhibit. A table headed "SPIRIT DATA CATEGORIES" carries a column titled "Google's Data Purchase Request," and each row is marked either Included or Not Included.

What is marked Included is the company's own working record. Microsoft 365 holds 100,000,000 emails, 500,000,000 Teams items, 17,082,644 OneDrive items and 20,577,677 SharePoint items, alongside 667,563 IT tickets, while operations contribute 763,391 flights flown and maintenance 787,452 parts received.

Business and engineering material sits on the same side. Revenue and pricing data, 175,658 employee records and 3,426,618 payroll records, 516 source code repositories holding roughly 30 million lines, and legal material including draft-to-final versioning histories of agreements are all marked Included.

The passenger-facing rows sit on the other side. Customer profiles covering 97,500,000 passengers, 50,200,000 Free Spirit members, 13,700,000 active email addresses, 30,865,471 call recordings, 15,784,473 chat sessions and the DOT complaint rows each carry the marking Not Included.

The proceeding began last year. The Spirit debtors filed their voluntary Chapter 11 petitions on August 29, 2025, and this auction was run under the bidding procedures order entered on June 22.

The purpose of the purchase comes from the reporting side. The Register writes that Google described the purchase as serving the improvement of its AI services, and that the company promised to strip out personal information wherever it is found.

OpenAI places 13- to 17-year-olds into ChatGPT for Teens automatically

OpenAI has introduced a mechanism that places users aged 13 to 17 into a separate experience called ChatGPT for Teens.

The announcement is dated August 18. The company describes it as an experience designed to help teens learn, think critically, deepen understanding, and use AI with confidence.

Two conditions trigger the placement. Where the system estimates that someone is under 18, or where the user states an age between 13 and 17, they are automatically placed into ChatGPT for Teens.

The learning side rests on four features. Study Mode works through guiding questions and step-by-step support so that a teen arrives at the material rather than receiving a finished answer.

Responsible homework reminders recognise when a teen appears to be shortcutting an assignment and redirect them toward Study Mode for collaborative problem solving.

Quizzes and Learning Visualizations give teens ways to practise, test what they know, and see difficult concepts more clearly.

Study Hours lets a teen or a parent choose the hours in which Study Mode is on by default, turning study habits into a routine.

Protections come on by default. For users under 18, the age-appropriate safeguards follow the Under-18 Principles in the company's Model Spec, with model-level and product-level interventions in higher-risk areas such as self-harm, violence, eating disorders, dangerous activities, and explicit sexual or graphic content.

The family tools widen as well. Parents with linked teen accounts can set Quiet Hours, manage selected settings, and receive safety notifications in limited high-risk situations; notifications related to eating disorders join that set here.

The shape of the relationship is drawn too. The updated under-18 model spec puts romantic language, encouragement of emotional dependence, and any suggestion of feelings or consciousness within its scope, and pairs that with break reminders and product cues that identify ChatGPT as AI.

The company frames the launch as an accumulation. It builds on parental controls, the Under-18 Model Spec, the Teen Safety Blueprint, and the rollout of age prediction.

An education partnership landed the same day. OpenAI announced a partnership with CodeAI focused on the education around the technology — helping teens understand how AI works, direct it, question it, and create with it.

Read the three together and the party drawing the line is the same party each time. The decision to stop, the decision to go and get a record, and the decision about where a user is sent all came from inside; the one place an outside procedure decides the outcome is the Spirit data.

Other Developments

Products

Ars Technica reported that researchers at the security firm Varonis, by putting questions to Microsoft 365 Copilot, coaxed out an undocumented parameter, "autorun=1," that makes the assistant execute commands from a single click. According to the outlet, clicking a link carrying that parameter sends passwords and other secrets, Base64-encoded, to a server the attacker controls; Microsoft addressed part of the issue in February, three months after the report, and has since shipped a more comprehensive fix.

Google began rolling out Gemini in Chrome on Android to all US Android users. Subscribers to the paid AI Pro and AI Ultra plans also get "auto browse," in which the browser carries out everyday tasks such as bookings and orders on the user's behalf, and Google says the design includes prompt-injection detection and confirmation steps for consequential actions.

Research

Google described roughly a decade of work on flood forecasting with AI. It provides river flood forecasts up to seven days ahead and urban flash flood forecasts 24 hours ahead across 150 countries covering more than two billion people, published a method called Groundsource for data-sparse regions in March 2026, and reports that in Nigeria anticipatory cash transfers based on the forecasts more than doubled income and cut food insecurity by 90 percent.

Policy

OpenAI announced an initiative to strengthen democratic oversight of AI used in national security. It offers government institutions practical tools, training and expertise, and the company frames its guiding principle as reinforcing democratic institutions and distributing power, with consequential decisions left to democratic processes.

Business

Etched, which builds silicon for AI inference, raised $700 million in a round led by Jane Street at a $21 billion valuation. Against $5 billion in December 2025 and $10.3 billion at its July 2026 Series C, the move over the past month works out to roughly 2.0 times, by this portal's own arithmetic.

Anthropic's annualized revenue run rate passed $65 billion at the end of July, TechCrunch reported, citing Bloomberg. The same piece sets that against $47 billion in May and $9 billion at the end of last year, adds a Financial Times report of $100 billion to $120 billion for the full year, and records that Anthropic has declined to comment.

Asana said it rebuilt an ageing test system with OpenAI's Codex, finishing in about two weeks for roughly $12,000 work it had estimated at five years and about $6 million. Up to four coding agents ran in parallel while engineers checked progress twice a day and reviewed every change by hand.

Source: Selected by the editors from the AI news inbox (48 items collected on August 19, 2026 — 11 primary, 37 secondary).