Digital Services Act: EU Puts ChatGPT Under Strict Scrutiny

EU subjects ChatGPT, Reddit, and Roblox to the strictest DSA level

The European Commission has placed ChatGPT, Reddit, and Roblox under the strictest rules of the Digital Services Act. ChatGPT was classified as a Very Large Online Search Engine (VLOSE), while Reddit and Roblox were classified as Very Large Online Platforms. The trigger is the threshold of 45 million average monthly users in the EU, which all three services have exceeded. At the end of March, OpenAI reported more than 159 million ChatGPT users in the EU, Reddit over 57 million, and Roblox 46.6 million active players. This makes ChatGPT the first AI chatbot with this classification; the Commission justifies its categorization as a search engine by stating that the service responds to queries and also searches the web in the process.

The three services have four months, until the end of December 2026, to comply with the additional obligations. This includes assessing and mitigating systemic risks: the dissemination of illegal content, effects on minors, on the physical and mental well-being of users, on fundamental rights, electoral processes, and public security. In addition, there are annual audits by independent auditors, the sharing of data with the Commission and authorities, and data access for researchers in certain cases. The Commission also gains investigative powers to examine the functioning of the services and their underlying systems.

Supervision is divided: For ChatGPT and Reddit, the Commission is working with the Irish regulatory authority Coimisiún na Meán, and for Roblox, with the Dutch Authority for Consumers and Markets. Henna Virkkunen, Executive Vice-President of the Commission for Technological Sovereignty, Security and Democracy, stated that the three services will be held to a higher standard of scrutiny and accountability in the future, and announced that more platforms will be designated as soon as they reach the threshold.

With these new designations, the Commission now lists 28 services as very large platforms and search engines, including offerings from Google, Apple, Microsoft, TikTok, X, Amazon, and Shein. Violations of the DSA can be punished with fines of up to six percent of global annual turnover, and in cases of serious and repeated violations, there is a risk of exclusion from the EU market. So far, penalties have been imposed on three companies, including Temu, AliExpress, and X, which was ordered to pay $138 million for transparency violations and other issues. Reddit stated that it has been preparing for the requirements for a long time and intends to coordinate with EU regulators; OpenAI declared its intention to meet the additional requirements → Engadget, Forbes Middle East, Techpresso

Synthszr Take: The real power play lies in the definition: Brussels decides that a chatbot with web access is a search engine, and just like that, the DSA’s sharpest instrument applies without a single procedure having to be initiated. OpenAI provided the number itself, 159 million monthly users of the search function, thereby holding the door open for the Commission to walk through. Four months until the end of November is a tight deadline for a company that has never written a systemic risk assessment, and six percent of global turnover is not something that can be downplayed in any financial plan. The obligation to release data to authorities and vetted researchers weighs more heavily than any fine: For the first time, an outsider will get a structured insight into how ChatGPT compiles answers and what gets left out in the process. Washington calls this extortion. The 28 services in the highest tier show: Whoever gets to set the categories no longer needs to issue bans.

OpenClaw 2.0 is here

OpenClaw released version 2.0 on August 30, internally designated as v2026.8.1, calling it the biggest release in its history. It contains over 16,000 Pull Requests from 933 contributors, 569 of whom are first-time contributors; this accounts for about half of all changes ever merged into the project. The numbers are self-reported but align with the activity in the public repository, which had approximately 388,000 stars and 81,000 forks at the time of release. This was preceded by an unusual pause: The team had previously shipped 106 versions in 230 days, usually one to two days apart, and then nothing for almost seven weeks.

Much less was planned. According to the release blog, the team only wanted to simplify the installation and rebuild the browser interface; both projects were tied to the rest of the system, and it turned into an overhaul of installation, messaging, memory, skills, model integration, automations, native apps, plugins, and security. The setup now utilizes what is already on the computer: existing ChatGPT or Claude subscriptions, API-Keys, and local models via Ollama or LM Studio, and it verifies that the selected model is actually responding before saving. Optional configuration has been moved out of the initial setup and can be configured later in conversation with the agent. The documentation describes a path from installation to the first browser conversation in about five minutes.

The browser app has been rebuilt from the ground up and opens directly into a conversation, with a sidebar, dockable panels for files and Git changes, and a side thread for intermediate questions that doesn’t interrupt the current task. Session transcripts are moved from files to an SQLite database, enabling full-text search across past conversations; those who want to revert to older versions must restore previously archived transcripts. In a test against a simulated gateway, the startup time decreased from around 1.6 seconds to 575 milliseconds, and the number of JavaScript requests dropped from 140 to 45. This measurement comes from the project itself and has not been independently reproduced.

The most striking new feature are shared cloud sessions: A second person can join or take over an agent's ongoing work without losing context. Sessions can run either on the local gateway, on paired own devices, or on rented disposable machines via the provisioning tool Crabbox, which supports AWS and Hetzner, among others; provider credentials remain on the gateway, according to the documentation. The same documentation states that the new role and sharing controls are 'not tenant isolation and not a security boundary'. A gateway forms a trust domain, and sandboxing remains disabled by default. The release follows analyses by Cisco Talos and Kaspersky, which attributed 36 percent of the skills in the ClawHub marketplace to Prompt Injection, as well as the CVE-2026-65105 vulnerability in NemoClaw’s local Ollama server, disclosed by Cyera.

Functionally, this brings OpenClaw close to Grok Bot, which SpaceXAI released into beta on August 11, and to ChatGPT Work, which gained event-driven tasks for Gmail, Slack, and GitHub on August 25. Both bundle the model, computing power, and interface into a managed subscription. OpenClaw leaves operation, maintenance, and security to the user, retaining in exchange the free choice of model provider and execution location. In its marketing, the project poked fun at the long wait itself: At least they beat GTA 6. → Moneycontrol, OpenClaw Blog, Digit, The Decoder, implicator, Cyber Security News, Trending Topics, RuntimeWire, RuntimeWire

Synthszr Take: The plan included two tasks: a simpler installation and a new browser app. What came out was a release with 16,000 Pull Requests, containing half the project’s history, because 933 volunteers produce work faster than a maintainer team can review and integrate it. The fact that its own documentation explicitly marks the new team roles as not being a security boundary is the logical consequence of a roadmap driven by contributions rather than an architectural decision. The next version must enable sandboxing by default, otherwise the 36 percent of compromised skills will decide OpenClaw’s fate for it.

Google turns its own e-book library into a source for Gemini Notebook

Google is integrating users' e-books as a source in Gemini Notebook, reports The Deep View. Purchased titles will thus land in the same workspace as uploaded PDFs and personal notes, and the assistant’s answers are intended to be based on these specific sources instead of general model knowledge (Grounding). The newsletter categorizes the move as an expansion of source-grounded AI. According to the editorial team’s assessment, the utility would increase significantly if the feature were extended beyond Google’s own library to include Kindle and other e-book providers. → The Deep View

Synthszr Take: For students, daily life is changing in a very specific way: The half-remembered passage from the seminar reader can be found in seconds, with a reference, instead of an evening of flipping through pages. The purchased library becomes a queryable work resource, and for anyone who works with long texts, this is worth more than the next point on some benchmark. Authors are on the other side of the same function because their book becomes context, from which someone can extract the core thesis without ever having read chapter three.

A stolen cookie beats any password: Claude quotas as new loot for infostealers

Anthropic warns that commercial malware is hijacking active Claude sessions and using them to drain the paid usage quotas of affected accounts. According to The Cyber Express, this involves infostealers, widely available malicious programs that read credentials, autofill data, and session cookies from browsers on infected computers. Anthropic identified six families, including Vidar, LummaC2, and RedLine on Windows, as well as Atomic Stealer on macOS; all of them are standard ware sold or rented on criminal marketplaces. With a valid session token, an attacker needs neither a password nor a second factor; they simply present themselves to the service as a logged-in user. Anthropic logs out affected accounts, removes stored payment methods, and refunds costs it deems unauthorized. → The Cyber Express

Synthszr Take: The choice of words does some quiet work: 'commodity infostealer' places the problem on the infected laptop and elegantly leaves open the question of long-lived, device-independent session tokens. The fact that Anthropic logs out accounts and reimburses damages is decent; however, the actual call to action still falls on the user, who is supposed to keep their computer clean. A once-stolen cookie remains a permanent pass as long as the session is not tied to a device or location, and this architectural decision lies solely with the provider. For a subscription with a monthly limit, this is an annoying bill; for accounts with agents that have API rights and file access, the depleted quota turns into a data breach.

Bank of England: Cybercrime could shake the financial system

Bank of England Governor Andrew Bailey, as head of the Financial Stability Board, sent a two-page letter to the G20 finance ministers and central bank governors just before their meeting in Asheville, North Carolina. The most urgent concern for the financial system: the impact of frontier AI on cyber risk. The models could change the speed, scale, and economics of attacks and undermine market confidence system-wide, especially since a few third-party providers are highly concentrated. Disruptions spread across jurisdictions via common technology providers and shared infrastructure. Bailey recommends preparing for simultaneous failures of multiple institutions and, if necessary, restoring critical systems from 'bare metal'. The letter also warns of the credit-financed wave of AI investment: leverage, stretched valuations, and market concentration make markets vulnerable to a disorderly correction across borders. The crash headlines thus pick up on a real part of the letter but narrow it down to the stock market aspect.

Synthszr Take: Bailey packs two risks into one letter: operational cyber-cascades across supervisory borders and a credit-leveraged valuation bubble. The headlines seize on the bubble because everyone immediately understands the word, while cross-border cyber resilience doesn’t excite anyone. Yet the cyber part is more concrete: Bare-metal recovery and faster patching as soon as AI models detect vulnerabilities more quickly are a to-do list that banks could start working on immediately. Anyone who only reads about the bubble misses the part of the letter with the call to action.

Clean product data becomes a revenue lever: Shopify delivers the first verifiable agent metric

Shopify is expanding its commerce infrastructure into a structured global product catalog that prepares merchant items in a machine-readable format for AI agents. In addition, there is the Universal Commerce Protocol, an open standard for transactions between AI agents and merchants developed jointly with Google, which is backed by Amazon, Mastercard, Meta, Microsoft, Walmart, and Visa, among others. Agents like ChatGPT and Copilot are intended to use it to find, recommend, and sell products directly. Shopify has also further developed its Sidekick tools for merchants. The core of the announcement is a number: according to Shopify, catalog-based AI searches convert at about twice the rate of searches based on scraped data. The company cites the complete, precise, and structured product data of the catalog as the reason. → MyClaw Newsletter

Synthszr Take: The doubled conversion rate is the first figure in all of agent-based commerce that a merchant can verify in their own dashboard. It says something inconvenient about data quality: if a cleanly structured catalog doubles the closing rate, every sloppily maintained product data set burns half of the agentic demand before it even reaches the shopping cart. Models recommend what they can reliably understand, and for most shops, variants, availabilities, and attributes are in a state that previously only a patient human could decipher. Data maintenance is thus moving from the back office to the income statement: anyone who cleans up their feed now is buying visibility before the competition even grasps the connection.

Claude Code: Anthropic sells a reduction as a gift

On September 14, Anthropic is permanently raising the standard weekly limits for Claude Code by 25 percent above the original baseline; this affects the Pro, Max, Team, and the seat-based Enterprise offerings. Until then, the temporary 50 percent increase, which started on May 13 and has since been extended several times, will remain in effect. Measured against the current quota, users will lose about 17 percent, which Anthropic has now confirmed itself: The original announcement thread was deleted after criticism and replaced with a clarification that openly names the reduction. → AlphaSignal

Synthszr Take: A usage limit is a price that doesn’t have to call itself a price, and Anthropic is shifting the rate here without touching the bill. The temporary 50 percent increase was a subsidized window in which teams calibrated their workflows to a cost level that was never planned. Four extensions over the summer effectively turned the promo into the new normal, and that’s precisely how every developer measures their work week. Starting September 14, they’ll get 17 percent less execution for the same amount of money, and this difference is the share of inference costs that the lab no longer wants to bear itself.

Product Managers ship SaaS without ever looking at the code

In his newsletter The Product Compass, Paweł Huryn shows how product managers can build a commercial web application without programming it themselves. Part 1 of the series guides you in three to four hours through the design, setup, security, and deployment of an application with multi-tenancy, built on Next.js, Clerk, and Supabase and controlled via Claude Code. The case study, AskOne, is a tool for live Q&A sessions at lectures and webinars: the audience types a join code into their phones and asks questions without an account; Huryn positions it as a five-dollar alternative to Slido. The process starts with a file called AGENTS.md, which captures the strategic context: market segments, value propositions, roles, and explicitly what is out of scope. The tutorial covers a multi-tenant setup, Google login, paid subscriptions, roles, and CI/CD, all without opening a single code file. As proof, Huryn cites his own project, Grok Build for VS Code: according to him, it has over 105,000 installations, 37,000 monthly active users, 5,094 unit tests, and around 177,000 lines of code, none of which he has read. → Paweł from The Product Compass

Synthszr Take: 177,000 lines of code, 5,094 tests, not a single file opened: This equation tangibly shifts the boundary between product management and development. The winners are the people who write down their intent so precisely that an agent can derive test cases and verification criteria from it; the AGENTS.md file with value propositions, roles, and non-goals thus becomes the actual work product. The intermediary roles come under pressure, i.e., the translators who turn requirements into task lists and task lists into status reports. What remains is the hard side of system, architecture, and security, and the numbers show where the problem lies: 37,000 monthly users versus $1,920 in annual revenue.

Apple Forensics: Ex-Engineer Allegedly Used Apple Circuit Diagram at OpenAI

Apple submitted a new piece of evidence on Monday in its trade secret theft case against OpenAI, basing it on the forensic analysis of a MacBook. The device belonged to Chang Liu, a former Senior System Electrical Engineer at Apple, who left the company for OpenAI in January 2026. According to Apple’s filing, Liu had kept the Apple device after his departure; his lawyers only handed it over on August 21 as part of the proceedings. Apple sued Liu and former Product Design Vice President Tang Tan in July, and in August, the company expanded the allegation to include a total of eleven former employees.

The filing makes four key claims. In March, two months after his departure, Liu allegedly downloaded dozens of confidential Apple files via a third-party cloud storage service, including the schematic of a voltage converter, and used it in his work at OpenAI. He and other individuals at OpenAI were aware of this access. After Liu learned of an internal investigation at Apple, he allegedly sent instructions to an OpenAI colleague on how to reset Apple devices, which she agreed to do. Furthermore, Liu allegedly used a tool at OpenAI that has the same name as an internal Apple development application.

Specifically, Liu is alleged to have used the circuit diagram in the simulation tool LTspice. In quoted messages from that time, he describes how his AI agent learned within an hour to operate LTspice, read results, and adjust compensation parameters; a task that previously took him a day was thus reduced to two hours. According to its own account, Apple only became aware of this because Liu used the file on a Mac mini, which later synchronized with the Apple MacBook via iCloud. The company is now also demanding access to this Mac mini and other devices in Liu’s possession.

Legally, Apple argues that feeding a trade secret into a learning system could create an irreversible and propagating use. The company calls the findings “shocking evidence” and is requesting an Expedited Discovery proceeding and an injunction, arguing that evidence could otherwise disappear. OpenAI has filed a motion to dismiss, calling the allegations baseless and publicly describing the lawsuit as “careless, aggressive and oddly personal”; it claims it neither wants nor needs to use others' trade secrets and is building products that are “entirely new.” According to industry reports, OpenAI has hired more than 400 employees from Apple. Judge Edward J. Davila is set to hear the arguments on October 1. → Reuters, 9to5Mac, AppleInsider, MacRumors, Engadget

Synthszr Take: A circuit diagram for a voltage converter is ultimately a file, but the knowledge of which compensation parameters to adjust is in the engineer’s head, and OpenAI has bought over 400 instances of exactly that from Apple. The quoted chat history shows the moment this knowledge changes form: Liu teaches an agent to operate LTspice, turning a day’s work into two hours. This turns experiential knowledge into training material, and Apple’s argument—that a learning system permanently perpetuates the use of a trade secret—is technically difficult to refute. For any company with hardware know-how, this means in practice: An exit interview and a locked account are not enough if a third-party cloud storage service still allows downloads two months after departure. On October 1, Judge Davila will hear the arguments, and regardless of the verdict, the question of who owns domain knowledge once it has passed through a model will rewrite the employment contracts of the coming years.