Privacy and data protection

Anthropic uncovers large-scale AI model theft operations

Three AI laboratories have been found conducting large-scale illicit campaigns to extract capabilities from Anthropic’s Claude AI, the company revealed.

DeepSeek, Moonshot, and MiniMax used around 24,000 fraudulent accounts to generate more than 16 million interactions, violating terms of service and regional access restrictions. The technique, called distillation, trains a weaker model on outputs from a stronger one, speeding AI development.

Distilled models obtained in this manner often lack critical safeguards, creating serious national security concerns. Without protections, these capabilities could be integrated into military, intelligence, surveillance, or cyber operations, potentially by authoritarian governments.

The attacks also undermine export controls designed to preserve the competitive edge of US AI technology and could give a misleading impression of foreign labs’ independent AI progress.

Each lab followed coordinated playbooks using proxy networks and large-scale automated prompts to target specific capabilities such as agentic reasoning, coding, and tool use.

Anthropic attributed the campaigns using request metadata, infrastructure indicators, and corroborating observations from industry partners. The investigation detailed how distillation attacks operate from data generation to model launch.

In response, Anthropic has strengthened detection systems, implemented stricter access controls, shared intelligence with other labs and authorities, and introduced countermeasures to reduce the effectiveness of illicit distillation.

The company emphasises that addressing these attacks will require coordinated action across the AI industry, cloud providers, and policymakers to protect frontier AI capabilities.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

OpenAI faces legal action in South Korea from top networks

South Korea’s leading terrestrial broadcasters have filed a lawsuit against OpenAI, claiming that the company trained its ChatGPT model using their news content without permission. KBS, MBC, and SBS are seeking an injunction to halt the alleged infringement and to recover damages.

The Korea Broadcasters Association said OpenAI generates significant revenue from its GPT services and has licensing agreements with media organisations worldwide.

Despite this, the company has refused to negotiate with the South Korean networks, leaving them without recourse to ensure proper use of their content.

The lawsuit emphasises the protection of intellectual property and creators’ rights, arguing that domestic copyright holders face high legal costs and barriers when confronting global technology companies. It also raises broader questions about South Korea’s data sovereignty in the age of AI.

Earlier action against Naver set a precedent for copyright enforcement in AI applications.

Although KBS subsequently partnered with Naver for AI-driven media solutions, the current case underscores continuing disputes over lawful access to broadcast content for generative AI training.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!  

Medical AI risks in Turkey highlight data bias and privacy challenges

Ankara is seeing growing debate over the risks and benefits of medical AI as experts warn that poorly governed systems could threaten patient safety.

Associate professor Agah Tugrul Korucu said AI offers meaningful potential for healthcare only when supported by rigorous ethical rules and strong oversight instead of rapid deployment without proper safeguards.

Korucu explained that data bias remains one of the most significant dangers because AI models learn directly from the information they receive. Underrepresented age groups, regions or social classes can distort outcomes and create systematic errors.

Turkey’s national health database e-Nabiz provides a strategic advantage, yet raw information cannot generate value unless it is processed correctly and supported by clear standards, quality controls and reliable terminology.

He added that inconsistent hospital records, labelling errors and privacy vulnerabilities can mislead AI systems and pose legal challenges. Strict anonymisation and secure analysis environments are needed to prevent harmful breaches.

Medical AI works best as a second eye in fields such as radiology and pathology, where systems can reduce workloads by flagging suspicious areas instead of leaving clinicians to assess every scan alone.

Korucu said physicians must remain final decision makers because automation bias could push patients towards unnecessary risks.

He expects genomic data combined with AI to transform personalised medicine over the coming decade, allowing faster diagnoses and accurate medication choices for rare conditions.

Priority development areas for Turkey include triage tools, intensive care early warning systems and chronic disease management. He noted that the long-term model will be the AI-assisted physician rather than a fully automated clinician.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!  

AWS warns of AI powered cybercrime

Amazon Web Services has revealed that a Russian-speaking threat actor used commercial AI tools to compromise more than 600 FortiGate firewalls across 55 countries. AWS described the campaign as an AI-powered assembly line for cybercrime.

According to AWS, the attacker relied on exposed management ports and weak single-factor credentials rather than exploiting software vulnerabilities. The campaign targeted FortiGate devices globally and focused on harvesting credentials and configuration data.

AWS said the potentially Russian group appeared unsophisticated but achieved scale through AI-assisted mass scanning and automation. When encountering stronger defences, the attackers reportedly shifted to easier targets rather than persist.

The company advised organisations using FortiGate appliances to secure management interfaces, change default credentials and enforce complex passwords. Amazon said it was not compromised during the campaign.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Stanford speech warns of AI tsunami

Senator Bernie Sanders has warned at Stanford University in California that the US is unprepared for the speed and scale of the AI revolution. Speaking in California alongside Congressman Ro Khanna, he called the moment one of the most dangerous in modern US history.

At Stanford University, Sanders urged a moratorium on the expansion of AI data centres to slow development while lawmakers catch up. He argued that the American public lacks a clear understanding of the economic and social impact ahead and that New York is already considering a pause.

Khanna, who represents Silicon Valley in California, rejected a complete moratorium but called for steering AI growth through renewable energy and water efficiency standards. He outlined principles to prevent wealth from being concentrated among a small group of tech billionaires.

Sanders also raised concerns in California about job losses and emotional reliance on AI, citing projections of widespread automation. He called for a national debate in the US over whether AI will benefit the public or deepen inequality.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Carrefour accelerates AI-enabled transformation to 2030, following Walmart’s strategic playbook

According to reporting by Diginomica, Carrefour, one of Europe’s largest retail groups, is accelerating the adoption of AI across its business as part of a strategic transformation aimed at 2030.

Inspired in part by the AI-driven overhaul undertaken by Walmart in the US, Carrefour’s initiative is intended to reshape its logistics, pricing, forecasting and store operations to become more data-driven, efficient and responsive to consumer trends.

Key elements of Carrefour’s AI focus include supply chain optimisation, dynamic pricing and promotions, customer engagement, and store and back-office automation.

First, using AI to predict demand, manage inventories and reduce waste across national and regional networks. Then, algorithms adjust pricing based on real-time data to improve competitiveness and margin performance.

Personalised offers and recommendations powered by machine learning work to enhance loyalty and user experience. Finally, AI tools streamline staffing, task allocation, and routine merchandising processes.

The transformation plan emphasises enterprise data strategy as a foundation, from consolidating disparate data sources to deploying machine learning models that inform business decisions in near-real time.

Carrefour executives view AI not just as a set of point solutions, but as core to future competitiveness, citing early gains in forecasting accuracy and reduced waste.

Carrefour’s approach is part of a broader retail AI arms race in which large grocers leverage scale and data to drive efficiency and customer centricity, with Walmart often cited as a pioneer whose playbook demonstrates the strategic value of enterprise-wide AI.

The report also notes challenges ahead, such as aligning organisational culture, ensuring data quality and addressing privacy concerns around personalised offers.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

OpenClaw exploits spark a major security alert

A wave of coordinated attacks has targeted OpenClaw, the autonomous AI framework that gained rapid popularity after its release in January.

Multiple hacking groups have taken advantage of severe vulnerabilities to steal API keys, extract persistent memory data, and push information-stealing malware instead of leaving the platform’s expanding user base unharmed.

Security analysts have linked more than 30,000 compromised instances to campaigns that intercept messages and deploy malicious payloads through channels such as Telegram.

Much of the damage stems from flaws such as the Remote Code Execution vulnerability CVE-2026-25253, supply chain poisoning, and exposed administrative interfaces. Early attacks centred on the ‘ClawHavoc’ campaign, which disguised malware as legitimate installation tools.

Users who downloaded these scripts inadvertently installed stealers capable of full compromise, enabling attackers to move laterally across enterprise systems instead of being confined to a single device.

Further incidents emerged on the OpenClaw marketplace, where backdoored ‘skills’ were published from accounts that appeared reliable. These updates executed remote commands that allowed attackers to siphon OAuth tokens, passwords, and API keys in real time.

A Shodan scan later identified more than 312,000 OpenClaw instances running on a default port with little or no protection, while honeypots recorded hostile activity within minutes of appearing online.

Security researchers argue that the surge in attacks marks a decisive moment for autonomous AI frameworks. As organisations experiment with agents capable of independent decision-making, the absence of security-by-design safeguards is creating opportunities for organised threat groups.

Flare’s advisory urges companies to secure credentials and isolate AI workloads instead of relying on default configurations that expose high-privilege systems to the internet.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!  

Claude Code Security by Anthropic aims to detect and patch complex vulnerabilities

Anthropic has introduced Claude Code Security, an AI-powered service that scans software codebases for vulnerabilities and recommends targeted fixes. Built into Claude Code, the capability is rolling out in a limited research preview for Enterprise and Team customers.

The tool analyses code beyond traditional rule-based scanners, examining data flows and component interactions to identify complex, high-severity vulnerabilities. Findings undergo multi-stage verification, receive severity and confidence ratings, and are presented in a dashboard for human review.

Anthropic said the system re-examines its own results to reduce false positives before surfacing them to analysts. Teams can prioritise remediation based on severity ratings and iterate on suggested patches within familiar development workflows.

Claude Code Security builds on more than a year of cybersecurity research. Using Claude Opus 4.6, Anthropic reported discovering more than 500 long-undetected bugs in open-source projects through testing and external partnerships.

The company said AI will increasingly be used to scan global codebases, warning that attackers and defenders alike are adopting advanced models. Open-source maintainers can apply for expedited access as Anthropic expands the preview.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

EU–US draft data pact allows automated decisions on travellers

A draft data-sharing agreement between the EU and the US Department of Homeland Security would allow automated decisions about European travellers to continue under certain conditions, despite attempts to tighten protections.

The text permits such decisions when authorised under domestic law and relies on safeguards that let individuals request human intervention instead of leaving outcomes entirely to algorithms.

A deal designed to preserve visa-free travel would require national authorities to grant access to biometric databases containing fingerprints and facial scans.

Negotiators are attempting to reconcile the framework with the General Data Protection Regulation, even though the draft states that the new rules would supplement and supersede earlier bilateral arrangements.

Sensitive information, including political views, trade union membership and biometric identifiers, could be transferred as long as protective conditions are applied.

EU countries face a deadline at the end of 2026 to conclude individual agreements, and failure to do so could result in suspension from the US Visa Waiver Program.

A separate clause keeps disputes firmly outside judicial scrutiny by requiring disagreements to be resolved through a Joint Committee instead of national or international courts.

The draft also restricts onward sharing, obliging US authorities to seek explicit consent before passing European-supplied data to third parties.

Further negotiations are expected, with the European Parliament’s Committee on Civil Liberties, Justice and Home Affairs preparing to hold a closed-door review of the talks.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!  

EU drops revised GDPR personal data definition amid regulatory pressure

Governments across the EU have withdrawn the revised definition of personal data from the GDPR omnibus package, softening earlier proposals that had prompted strong resistance from regulators and civil society.

A decision that signals a preference for maintaining the original scope of the General Data Protection Regulation instead of reopening sensitive debates that risked weakening long-standing protections.

Greater attention is now placed on the forthcoming pseudonymisation guidelines prepared by the European Data Protection Board. These guidelines are expected to shape how organisations interpret key safeguards, offering practical direction instead of altering the legal definition of personal data.

The updated prominence given to the guidance reflects a broader trend within the Council towards regulatory clarity rather than legislative redesign.

The compromise text also maintains links with the wider review of the ePrivacy Directive, keeping future updates aligned with existing digital-rights rules.

Member states appear increasingly cautious about reopening foundational privacy concepts, opting to strengthen enforcement through guidance and implementation rather than altering core definitions in law.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!  

Diplo chatbot can make mistakes. Consider checking important information.