Privacy
Norway orders SATS to fix gym photo checks; Europol plan draws fire
Norwegian DPA forces SATS to change check-in photo legal basis, while EDPS warns Europol overhaul risks biometric rights.
This edition was produced with artificial intelligence. Text and voice are generated automatically.
Norway orders SATS to change check-in photo legal basis and honor objections
The Norwegian Data Protection Authority ordered SATS ASA on July 15, 2026 to change how it justifies the photographs it takes of gym members at check-in, ruling that the contractual necessity basis the chain published was incorrect and that members have a right to object. The decision, reference 25/05417-8, followed 26 complaints lodged between May and November 2025, including 19 forwarded by Finland. SATS has until September 11, 2026 to comply and document its changes; no fine was imposed, but the company was reprimanded on three counts and ordered to make four changes.
The authority concluded that photographing members at check-in cannot rest on Article 6(1)(b) of the GDPR and must instead rest on Article 6(1)(f), legitimate interests. That reclassification activates the Article 21 right to object, which SATS had told members did not apply to them. SATS has required check-in photos since 2004 and from August 11, 2025 planned to deny entry to members without a photo on file. The authority found the contractual basis test was not met, noting that primary gym access can be achieved by scanning a membership card, and that SATS already permits exemptions for police officers and individuals in sensitive professions. The decision does not conclude SATS actually meets the legitimate interest conditions; under Article 5(2), the company must document its assessment.
Norway forces SATS to let gym members object to check-in photos →
EDPS warns Europol overhaul would put non-suspect biometric data in cross-border database
The European Data Protection Supervisor formally declared in Opinion 18/2026, published on August 13, 2026, that the European Commission’s proposal to rebuild Europol fails to protect the biometric data of people with no established criminal connection. The opinion concluded the proposal would allow Europol to process photographs, fingerprints, and other biometric identifiers on such individuals for an unspecified and potentially indefinite period without adequate independent oversight. Supervisor Wojciech Wiewiórowski stated that fundamental rights cannot be sacrificed in pursuit of security.
The June 24, 2026 proposal would establish three technical systems, including a real-time biometric query environment and a Police Shared Data Space granting private entities like cloud providers access to Europol’s analytical environment. The budget would double to €3 billion between 2028 and 2034, and Europol’s staff would also double. The EDPS identified three concerns: lack of clear necessity and proportionality criteria, no defined data retention periods, and a self-authorization bypass allowing Europol to proceed with sensitive processing without prior EDPS approval. The opinion follows an ongoing legal conflict over Europol’s retention of uncategorized datasets, with an appeal at the Court of Justice still pending. Access Now and European Digital Rights criticized the reform as rewarding misconduct and selling fundamental rights to surveillance companies.
Europol Expansion Would Put Non-Suspect Fingerprints Into Cross-Border Police Database →
London police apologize after email addresses of Al-Fayed abuse complainants exposed
London police apologized Saturday after email addresses of complainants alleging sexual abuse by the late Harrods owner Mohamed Al-Fayed were accidentally shared with those on a communications distribution list containing victims and others. The breach occurred Tuesday when the Metropolitan Police sent a monthly update to victims but failed to hide recipients’ email addresses, leaving them visible to others on the list. Around 140 women were impacted by the error, according to BBC News, with U.K. media reporting that the data watchdog had been notified.
The Met said the issue was identified quickly and immediate action was taken, with affected individuals contacted directly on the day of the incident. The incident is being investigated as a matter of priority. To date, some 157 victims have reported allegations of sexual assault, rape, sexual exploitation, and human trafficking in relation to Fayed, who died aged 94 in 2023. The Met has so far questioned six people suspected of enabling Fayed’s offences, although it has not yet made any arrests.
Google’s HEIR brings homomorphic encryption to private AI inference
Google introduced HEIR, an open-source compiler toolchain and development platform for homomorphic encryption, as part of its Private Computing Toolkit. Homomorphic encryption allows computations to be performed directly on encrypted data, enabling servers to process ciphertexts and return encrypted results without exposing underlying information. This addresses the trade-off of standard protections like end-to-end encryption, where user data is protected from breaches but the service provider cannot offer features that depend on the data, such as spam or virus detection. HEIR can convert pre-trained AI models that operate on unencrypted data to operate on encrypted inputs.
Google’s stated vision is to make HEIR a one-click solution for non-experts to incorporate encrypted inference into production applications. Google has partnered with companies developing hardware accelerators for homomorphic encryption, including Belfort, Niobium, Cornami, and Optalysys, and plans to demonstrate latency benefits. HEIR has also become a research platform with collaborations involving Georgia Tech, Carnegie Mellon, and other universities. Google shared four private inference applications compiled with HEIR, including a deep learning recommendation model, a credit card fraud detector, an anomaly detection system for encrypted network traffic, and a hotword detection model.
How Google is Making Private AI Practical with Homomorphic Encryption →
AI text watermarks can be stripped by asking a second AI to paraphrase
Text watermarking hides a signature in the words a model chooses; the model favors words that score highly on a simple formula, and a high total score across a block of text indicates machine authorship. This watermarking is easy to remove: asking any weak, unwatermarked model to paraphrase the text strips it. Goedecke said that because the watermark is inherent to subtle vocabulary choices, re-wording the content will remove the watermark.
The EU AI Act requires watermark methods to be interoperable, meaning AI companies must publish how their watermarks work. This raises questions about the effectiveness of current watermarking approaches for AI-generated content detection.
You can remove an AI watermark from text by asking a second AI →
Consumer Reports flags security flaws in five home security cameras
Consumer Reports testing identified five home security cameras as poorly rated, with security vulnerabilities and data privacy concerns being the primary issues for most. The Arlo Essential Indoor Camera Plug-in VMC3070 has a significant undisclosed vulnerability making it susceptible to hacking, combined with poor overall data security and lack of transparency over data privacy. The Wyze Cam Floodlight Pro was poorly rated due to inadequate security and a serious undisclosed vulnerability, consistent with Wyze’s past security problems. The Philips Secure Battery Camera failed CR’s security testing due to a high-risk vulnerability that attackers could exploit.
The Arlo Ultra 3rd Gen VMS5250 was rated poorly not for security, but for poor video quality with soft images and unusable nighttime captures, plus slow detection alerts. The Wyze Cam Floodlight v2 has reasonable video quality but poor data security and a potentially exploitable vulnerability. CR suggested alternatives for each model, including TP-Link Tapo, Eufy, and Blink cameras.
5 Of The Worst Rated Home Security Cameras, According To Consumer Reports →
CRYPTO 2026 opens with record papers, post-quantum scrutiny, and Encrochat analysis
CRYPTO 2026, the 46th edition of the International Association for Cryptologic Research’s flagship conference, opens its main program Monday at the University of California, Santa Barbara, running through Thursday, August 20. It is the first CRYPTO held since NIST finalized its post-quantum cryptography standards in August 2024, meaning the lattice mathematics now deployed on billions of devices will face scrutiny from researchers. The main program features 189 full papers selected from 781 submissions, published across ten Springer LNCS volumes—a record in the conference’s 46-year history.
Volume III contains nine cryptanalysis papers targeting post-quantum foundations, including an improved quantum algorithm for 3-tuple lattice sieving and a paper quantifying the security gap between standard LWE and the module variant used in ML-KEM. The program includes a new track on neural network cryptanalysis and machine learning, with papers applying algebraic attack methodology to convolutional neural networks and formalizing model extraction as a cryptanalytic problem. Volume IX, with 20 papers on zero-knowledge proofs, is the largest single-subject block. A paper on the Encrochat law-enforcement hack identifies the operation with the model of a covert adversary, while other highlights include robust single-trace key extraction and a formal security proof of masking.
CRYPTO 2026 Opens Today: AI Systems Are Now Both Tool and Target of Math Attacks →
Singapore opens first biological data centre using living human brain cells
Singapore’s first biological data centre, located inside the National University of Singapore’s Centre for Life Sciences, processes data using living human brain cells instead of silicon chips. Lab technicians feed the cells a cocktail of sugar, micronutrients, and pH buffers every three days, while a gas mixer supplies carbon dioxide, oxygen, and nitrogen as a life support system. The facility was developed by Australian biotech startup Cortical Labs and went live on July 16, with plans for up to 1,000 CL1 biological computers at the site pending regulatory approval.
The Singapore facility comprises 20 units of biological computers called CL1s, each housing at least 200,000 lab-grown brain neurons on an electrode-fitted silicon chip. Customers pay US$2,200 a month per CL1, about half the US$4,300 monthly fee major cloud platforms charge for renting a high-end AI chip. Each CL1 uses 30 watts, less than a handheld calculator, while an Nvidia H100 SXM chip can consume up to 700 watts. Cortical Labs founder Chong Hon Weng said biological data centres suit applications with limited datasets and highly unpredictable conditions, citing humanoid robots and cybersecurity anomaly detection.
Forget silicon chip servers, Singapore’s newest data centre needs to be fed →
More than $2 trillion in crypto assets at risk from quantum computing, industry urged to migrate
More than $2 trillion in digital assets—nearly the entire crypto market, valued at $2.16 trillion—is at risk from quantum computing, prompting calls for an industry-wide migration to quantum-resistant cryptography. Christopher Smith, co-founder and CEO of Quantus, a quantum-secure blockchain network, said the great quantum migration will require the entire digital asset industry to participate. Smith said over $2 trillion in digital assets is secured by elliptic curve cryptography, which has been known to be quantum-vulnerable for over 30 years.
Smith and Quantus identified Binance’s Bitcoin cold wallet, containing more than $10 billion, as an obvious attack target, and noted the administrative key controlling USDT could be even more dangerous. Coinbase cautioned against treating the entire crypto ecosystem as equally exposed, telling Fortune that bitcoin’s core infrastructure is largely safe and that the real vulnerability is at the wallet level. Google proposed a 2029 target for cryptocurrency systems to migrate away from vulnerable cryptography. Building a machine capable of executing a quantum attack remains beyond current capabilities, but Smith said being a year too early is much better than being a day too late.
Hong Kong firm Antimatter bets on Chinese open-weight models to rival CoreWeave
Antimatter, a Hong Kong-based neo-cloud provider established in May, is helping customers switch from dominant US AI models and cloud systems to Chinese open-weight models, aiming to cut costs and increase data control. The company expects revenue to grow exponentially over the next two years, arguing closed frontier AI models will not be able to satisfy rising global demand. Queenie Chan, Antimatter’s chief marketing officer, said many customers are seeking alternatives to frontier models, with quite a bit of inquiries coming in from the Middle East and Europe.
Antimatter cited a French accounting firm that had spent €50,000 monthly on US hyperscaler compute for AI tasks; after being acquired, the firm wanted data processed and stored in Europe. Antimatter deployed two Qwen models on French infrastructure, cutting the firm’s AI costs by 40 per cent. By end of 2024, Antimatter aims to deploy 100 Policloud data centres with more than 40,000 GPUs; by 2030, it targets 1,000 Policlouds with more than 400,000 GPUs. It expects revenue to reach €1 billion in 2028 and is raising €300 million in equity financing.
Hong Kong firm bets on Chinese open-weight models to rival CoreWeave →
Scientists warn of growing dependence on black-box tools they cannot control or fully understand
A new study by an international team of scientists warns that researchers are increasingly dependent on powerful data sources and tools—such as artificial intelligence, satellite imagery, online data, and digital sensors—that they cannot fully understand, inspect, or verify. These systems operate as scientific black boxes and can challenge trust in science. Lead author Ivan Jarić, a researcher at the University of Paris-Saclay, said many of these tools keep the processes behind their results largely hidden, often owned by private companies that intentionally limit access to information about how their systems operate.
The paper identifies several types of black boxes used in ecology and conservation, including large language models, proprietary remote sensing products, wildlife tracking devices that withhold raw data, and online platforms with hidden algorithms. The growing dependence is strengthened by a publish-or-perish culture and pressure to increase productivity. The authors recommend prioritizing open-source software and hardware, benchmarking proprietary tools against transparent datasets, and keeping human oversight central throughout the research process. The study is titled The black-box future of ecology and conservation and was published in the journal BioScience.
New Jersey data broker law enforcement delayed after political concerns emerge
A dispute is escalating in Trenton after Assemblyman Brian Bergen accused Attorney General Jennifer Davenport and Gov. Mikie Sherrill’s administration of unlawfully refusing to enforce a newly enacted data broker law days after it was signed. The controversy centers on A-5328, signed into law on June 30, which creates one of the nation’s strictest regulatory frameworks for data brokers. Within days, the state announced it would delay enforcement of major portions of the law while lawmakers consider technical changes.
Bergen, a Republican representing New Jersey’s 26th Legislative District, argues the executive branch lacks the constitutional authority to suspend enforcement of a law. The law broadly regulates data brokers but did not expressly exempt political campaigns or vendors that maintain voter-targeting databases. Campaign consultants warned the law could disrupt access to voter data relied upon during elections. The Sherrill administration said portions of the law contain defects requiring legislative correction before enforcement proceeds, with the Office of Consumer Protection announcing it will launch the required registry in spring 2027.
Davis County approves license plate camera deal amid privacy backlash
Davis County commissioners voted 2-1 on Tuesday to approve a memorandum of understanding governing how the county’s automatic license plate reader cameras share data with the Utah Department of Public Safety, over the objections of residents who packed the meeting to voice fears about privacy and out-of-state surveillance. Commissioners Bob Stevenson and Lorene Kamalu voted in favor, while Commissioner John Crofts cast the lone dissenting vote. The five-year agreement restricts use of license plate data strictly to law enforcement purposes and requires it to be stored in a secure government cloud.
The vote came two days after public records showed that 96.6% of the roughly 5.1 million searches run on Weber County’s 10 Flock cameras between February 2022 and July 2026 originated from out-of-state law enforcement agencies, with Houston Police Department alone accounting for more than a third of all searches. Weber County Sheriff’s Office spokesperson Colby Ryan said monthly internal audits confirmed the out-of-state queries were manual, officer-initiated searches backed by documented justifications. The Utah State Privacy Commission is currently conducting a formal study on law enforcement use of the technology, with official policy recommendations expected by the end of September. Efforts to set statewide ALPR standards through the Legislature have repeatedly failed.
Davis County Splits 2-1 on License Plate Camera Deal Amid Privacy Backlash →
WhatsApp tests AI-powered Scam Alert feature to flag suspicious messages
WhatsApp, which has more than three billion monthly active users worldwide, is testing a new optional feature called Scam Alert. The tool runs an on-device machine learning model to flag potential scam messages and is currently rolled out on a limited scale before becoming available to all users. The feature is being developed partly in response to concerns that new WhatsApp usernames, introduced in late June to protect phone-number privacy, could be used by scammers to impersonate people, brands, or public figures.
Once a user enables Scam Alert, it downloads a machine learning model to the device that classifies whether incoming messages from non-contacts match known scam patterns. The model is trained on patterns from scam conversations reported by users to Meta and performs probabilistic classification based on conversational structure and linguistic signals. Meta stated that no content is automatically reported to WhatsApp, Meta, or any third party. If the model identifies a likely scam attempt, the user sees a warning in the chat that is not visible to the other person. The model and message data remain on the device, with all inference happening on-device and no message content leaving the device for classification.
Scam Alert on WhatsApp: Meta tests new AI-powered online safety tool →