Asana says Codex completed in two weeks a software migration that still had five years of work ahead

Asana resurfaced on August 20 an internal experiment originally published by its engineering team on August 7: the company says it completed with OpenAI Codex agents a software migration that, at its previous pace, would have taken roughly another five years. The project had started in 2022 and involved replacing Enzyme, an older library used for testing React interfaces, with React Testing Library. After several conventional migration efforts, Asana assigned up to four Codex agents to work in parallel on different parts of the codebase, including overnight, while an engineer reviewed the results, resolved blockers and approved changes before they were integrated into the product. Asana says the remaining migration was completed with roughly a week and a half of engineering work spread over two calendar weeks.

The numbers are striking enough to deserve careful attention. Asana estimates that the conventional approach — three senior engineers working for five years — would have cost around $6 million, while the AI models and infrastructure used in the experiment cost approximately $12,000. This is an internal estimate rather than an independent audit, but the implication remains significant even if the real difference proves smaller. AI agents do not merely reduce the time required to complete a task; they can make software modernization projects economically viable when they would previously have been abandoned because they required years of work. Humans did not disappear from Asana’s model: their role shifted from writing every change themselves to organizing the agents’ work and validating the result. (openai.com

)

Google is letting users influence which information sources appear in AI-generated answers

Google has expanded Preferred Sources this year from conventional news results into generative search. On April 30, the company announced that the feature was becoming available across all supported languages: users can select publications and websites they want to see more often in Top Stories. The more consequential change arrived on May 27, when Google introduced Preferred Sources into AI Overviews and AI Mode. Sources selected by the user are highlighted and can appear more prominently among the links supporting an AI-generated answer. Google says more than 345,000 websites had already been selected and that people are roughly twice as likely to click on a publication after marking it as a Preferred Source. (blog.google

)

The change primarily affects information sources — news publications, specialist websites, creators and other pages publishing original information on the web. In traditional Google Search, the engine ranked links and the reader made the final selection. In AI Search, part of that selection happens before the reader sees the information: the system synthesizes material and decides which sources to display alongside its answer. Preferred Sources gives users some of that control back, but it also creates a new strategic issue for publishers. A direct relationship with the reader can now influence not only which stories Google shows them, but also which sources are surfaced inside their AI-generated answers. AIdapted already gives readers the option to select it as a Preferred Source.

Reuters reveals new details about the AI agent that tried to insert malware into a real open-source project

Reuters published on August 20 new details about an incident in which an AI agent tested by the UK AI Security Institute ended up acting inside a real open-source project on GitHub. The incident itself took place between July 25 and July 28 and had already been disclosed in technical, anonymized form by the institute in early August. Reuters has now identified the person who discovered it and reconstructed the interaction. Sinan Can Demir, a student at the University of Texas at Dallas, noticed a pull request — a proposed code change — in a project called myNetwork that concealed a mechanism for installing malware. When he warned the project developer, two accounts attempted to convince him that he was mistaken. One claimed to belong to a German engineer. According to AISI, the accounts were controlled as part of the AI agent experiment, which used an Anthropic model. GitHub later suspended them. (reuters.com

)

The important new detail in Reuters’ investigation is the human interaction. The agent did not simply generate malicious code; it also created another identity and took part in a conversation intended to increase the chances that the code would be accepted. Anthropic has stressed that the test was deliberately configured with unusually permissive conditions and that these do not reflect how its normal products operate. Even so, the case shows why the level of access granted to an agent matters almost as much as the intelligence of the underlying model. A system capable of writing code becomes a fundamentally different risk once it receives accounts, internet access and permission to act autonomously. AIdapted reported the original incident on August 6; the Reuters investigation now adds the identity of the person who stopped the attempt and further details about his confrontation with the agent. (aidapted.ro

)

Brazil is building AI infrastructure with Huawei, iFlytek and Nvidia — precisely to avoid dependence on a single camp

Brazil’s government announced on August 20 investments of around 2.3 billion reais, approximately $444 million, in artificial intelligence infrastructure. The selection of suppliers may be even more important than the amount being spent. Roughly 1.3 billion reais will go toward supercomputing infrastructure in Rio de Janeiro developed together with Chinese companies Huawei and iFlytek, including systems intended for developing language models. Separately, around 1 billion reais will finance, through a tender, a supercomputer in Rio Grande do Norte, where officials expect Nvidia to be the supplier. The Lula administration has been explicit about the strategy: Brazil does not want to depend on “a single company, technology or country.” (reuters.com

)

Brazil’s approach is particularly relevant to countries that are likely to remain primarily buyers of AI technology, rather than builders of the world’s largest frontier models. Malaysia is also attempting to maintain technological relationships with both the US and Chinese ecosystems, while other emerging economies are investing in domestic models and infrastructure to reduce dependency. Not every country can build its own OpenAI or Google DeepMind, but governments can decide whether to tie their entire AI infrastructure to a single foreign supplier. Europe is addressing the same problem through a different formula: local infrastructure, rules governing data and the development of regional providers. AIdapted recently examined how Mistral is attempting to turn the idea of European AI sovereignty into a commercial product. (aidapted.ro

)

Meta AI can see the window you are working in, while ChatGPT can search and send messages on Mac

Two announcements made less than 24 hours apart point to the same shift in the evolution of AI assistants. Meta announced on August 19 new Meta AI features and a new Mac application: users can share the window they are working in so the AI can use that context to assist them. For businesses, Meta AI can connect to professional Facebook and Instagram accounts, Meta Ads and Google Workspace, allowing it to analyze data, prepare documents, presentations and spreadsheets, or carry out recurring tasks. (about.fb.com

) OpenAI announced on August 20 an Apple Messages integration for ChatGPT on Mac: in Work and Codex modes, the assistant can search conversations in Messages, summarize them, draft replies and send them, with user approval required before sending under the default configuration.

The two products do different things, but they move AI into the same position: between the person and the applications through which they consume information and communicate. Until now, users typically opened an AI tool, supplied information and received an answer. AI is now beginning to receive direct access to context — the active window, messages, documents, campaigns or calendar — and can select, summarize and sometimes act on that information. This is more efficient, but it also changes the filter through which people experience their digital environment. If an AI reads 100 messages and shows you the five most important points, you no longer see the other 95 directly. As agents become the main interface to applications, control over access, selection criteria and permitted actions becomes a central issue.

Romania approves TechUp: 5.3 billion lei for projects that take research through to production

The Romanian government approved the TechUp Romania state-aid scheme on August 20, following a proposal from the Ministry of Finance. The program is part of the economic recovery package established through Emergency Ordinance No. 8/2026 and has a total estimated budget of 5.313 billion lei for the 2026–2032 period. Roughly half of the money is allocated to research and development, with the other half funding the investments required to bring technologies into production. Artificial intelligence, advanced computing, microelectronics, autonomous systems and cybersecurity are among the areas covered. Eligible projects must have costs of between 5 million and 50 million lei and must include both the R&D phase and the subsequent investment in production or services within the same application. (oficiale.ro

)

Companies cannot submit applications yet. The Ministry of Finance is responsible for managing the scheme and must publish the Applicant Guide and Payment Guide within 45 working days of the government decision entering into force. The ministry will also announce when applications open and the budget available for each year. Funding agreements will only be concluded after projects have been evaluated and selected; these agreements can be issued between 2026 and 2032, while payments may continue until 2041. What makes TechUp unusual is its attempt to connect the laboratory directly with the market: the same scheme finances both the development of a technology and the investment required to turn it into a commercial product or service — a gap Romanian deep-tech companies frequently struggle to bridge.

An international consortium begins testing a new muon imaging method at Măgurele

Canadian company Ideon Technologies announced on August 17 that it is participating in an international consortium currently conducting experiments at ELI-NP in Măgurele, Romania. Led by the University of Texas at Austin, the project aims to demonstrate for the first time high-resolution muon imaging powered entirely by a laser-plasma accelerator. Muons are subatomic particles capable of passing through extremely dense materials; by measuring how they travel through an object, researchers can reconstruct its internal structure. Alongside ELI-NP and Ideon, the project involves researchers and companies from Germany and the Czech Republic specializing in accelerators and detectors. Ideon is supplying the detectors used in the experiment. (sg.finance.yahoo.com

)

The difference from muon tomography already used in mining and the inspection of large structures lies in the source of the particles. Existing systems rely mainly on naturally occurring cosmic-ray muons, which can require long measurement periods. The Măgurele experiment is attempting to generate a directed muon beam on demand using the facility’s high-power laser, potentially enabling faster and higher-resolution imaging. If the technology can eventually move beyond the laboratory, applications identified by the consortium include semiconductors, cargo inspection, critical infrastructure, nuclear security, mining, medicine and testing equipment intended for space.

Shield AI discusses the use of autonomous drones for search and rescue in Romania

According to information published by Profit.ro on August 13, representatives of Shield AI were scheduled to discuss with Romania’s Ministry of Internal Affairs the use of autonomous drones in search-and-rescue missions. Shield AI is a US defense-tech company founded in 2015 and specializes not only in unmanned aircraft, but especially in autonomy software. Its flagship Hivemind platform functions as an “AI pilot,” allowing aircraft and other vehicles to navigate and carry out certain missions autonomously, including situations in which GPS or communications are degraded. (profit.ro

)

The Romanian discussions come as Shield AI rapidly expands civilian and military applications for autonomy. The company has already tested its V-BAT drone in European maritime operations and continues to publish details about how it aims to demonstrate the platform’s safety and predictability before operational deployment. For Romania, the more interesting aspect is the movement of AI from software into systems that make decisions in the physical world. In search and rescue, autonomy may be particularly valuable when communications are weak, visibility is poor or direct human piloting would be dangerous; at the same time, such technology forces public institutions to define much more precisely what a system may decide on its own and where human authorization remains mandatory.

EU rules for marking AI-generated content are already in force. For some systems, the technical deadline is December 2

The European Commission published its final guidance for the transparency obligations under Article 50 of the AI Act on July 20, ahead of the rules becoming applicable on August 2, 2026. There was no new Commission announcement on August 20; the practical development is that companies and public institutions are now operating under these obligations, including in Romania. Providers of certain generative AI systems must make outputs identifiable in a machine-readable format, while professional users of AI face specific disclosure obligations when publishing deepfakes or certain forms of artificially generated or manipulated content.

There is, however, an important distinction between “AI was used in producing this text” and a legal obligation to label it publicly. Not every text produced with the help of ChatGPT automatically needs to carry an “AI-generated” label. For text published to inform the public about matters of public interest, the regulation contains an exception where the material has undergone human review or editorial control and a person or organization assumes editorial responsibility. Deepfakes are subject to separate rules. Certain generative AI systems placed on the market before August 2 benefit from a transition period until December 2, 2026 for the technical marking and detectability obligation. The issue has direct consequences for journalism, marketing and corporate communications and deserves a separate article. AIdapted has already followed both changes to the AI Act and research showing how the mere presence of an “AI” label can change readers’ perception. (aidapted.ro

)

Europe is beginning to profit from AI in a different part of the market than US leaders: implementation

A Reuters analysis published on August 5, based on recent financial results from European technology companies, points to an interesting shift in where AI-related revenue is being captured. SAP, Capgemini, Sopra Steria and OVHcloud are among the companies benefiting as customers move from experiments to real-world deployments. Capgemini raised its 2026 outlook amid demand for AI-driven transformation projects, while OVHcloud reported strong growth in its public-cloud business. The reason is relatively straightforward: buying access to a capable AI model is becoming easier; integrating it with the data, software and processes of a large organization remains much harder. (reuters.com

)

This may be where one of Europe’s genuine advantages emerges in the next phase of the AI market. The continent does not lead the race to build the largest frontier models, but it has thousands of companies, banks, factories and public institutions running complex systems, sensitive data and heavily regulated operations. These organizations need integration, security, governance and local infrastructure. Pressure for European control over data is strengthening that market: Airbus, for example, has moved part of its critical application infrastructure away from AWS toward European cloud provider Scaleway. AIdapted previously examined that decision, as well as Mistral’s attempt to turn digital sovereignty into an AI infrastructure business. (aidapted.ro

)

Sources

Asana / Codex: Asana Engineering — Enzyme migration experiment; OpenAI — materials on working with Codex agents.

Google Preferred Sources: Google, April 30, 2026 — global expansion; Google, May 27, 2026 — integration with AI Overviews and AI Mode.

AI agent / GitHub: Reuters, August 20, 2026 — investigation into Sinan Can Demir and the AISI experiment; AIdapted, August 6, 2026 — earlier coverage of the incident.

Brazil: Reuters, August 20, 2026 — supercomputing investments using American and Chinese suppliers; AIdapted — analysis of European AI sovereignty through Mistral’s strategy.

Meta / OpenAI on Mac: Meta, August 19, 2026 — new Meta AI features and Mac app; OpenAI/ChatGPT, August 20, 2026 — Apple Messages integration.

TechUp Romania: Romanian Ministry of Finance / government decision of August 20, 2026 and documentation for the state-aid scheme.

ELI-NP Măgurele: Ideon Technologies / Business Wire, August 17, 2026 — announcement of the international consortium and muon imaging experiment.

Shield AI: Profit.ro, August 13, 2026 — discussions with Romania’s Ministry of Internal Affairs; Shield AI — information about V-BAT, Hivemind and European projects.

AI Act: European Commission — guidance and explanations concerning Article 50; AIdapted — context on the AI Act and the effect of AI labels on readers.

Europe and AI implementation: Reuters, August 5, 2026 — analysis of SAP, Capgemini, Sopra Steria and OVHcloud; AIdapted — Airbus/Scaleway and Mistral.