Network security

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! 

AI data centre surge pushes electricity demand in the UK to new heights

The UK faces rising pressure on its electricity system as about 140 new data centre projects could demand more power than the country’s current peak consumption, according to Ofgem.

The regulator said developers are seeking about 50 gigawatts of capacity, a level driven by rapid growth in AI and far beyond earlier forecasts.

Connection requests have surged since late 2024, placing strain on a grid already struggling to support vital renewable projects that are key to national climate targets.

Work needed to connect expanding data centre capacity could delay schemes considered essential for decarbonisation and economic growth, instead of supporting the transition at the required pace.

The growing electricity footprint of AI infrastructure also threatens the aim of creating a virtually carbon-free power system by 2030, particularly as high costs and slow grid integration continue to hinder progress.

A proposed data centre in Lincolnshire has already raised concerns by projecting emissions greater than those of several international airports combined.

Ofgem now warns that speculative grid applications are blocking more viable projects, including those tied to government AI growth zones.

The regulator is considering more stringent financial requirements and new fees for access to grid connections, arguing that developers may need to build their own routes to the network rather than rely entirely on existing infrastructure.

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! 

Wikipedia removes Archive.today links

Wikipedia editors have voted to remove all links to Archive.today, citing allegations that the web archive was involved in a distributed denial of service attack.

Editors said Archive.today, which also operates under domains such as archive.is and archive.ph, should not be linked because it allegedly used visitors’ browsers to target blogger Jani Patokallio. The site has also been accused of altering archived pages, raising concerns about reliability.

Archive.today had previously been blacklisted in 2013 before being reinstated in 2016. Wikipedia’s latest guidance calls for replacing Archive.today links with original sources or alternative archives such as the Wayback Machine.

The apparent owner of Archive.today denied wrongdoing in posts linked from the site and suggested the controversy had been exaggerated. Wikipedia editors nevertheless concluded that readers should not be directed to a service facing such allegations.

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!  

Data breach at PayPal prompts password resets and transaction refunds

PayPal has notified some customers of a data breach linked to its Working Capital loan application, after unauthorised access between 1 July and 12 December 2025 exposed personal information. Letters dated 10 February confirm that around 100 customers were potentially affected.

The incident was linked to an error in the Working Capital application, described as a ‘code change’. PayPal said it ‘terminated the unauthorised access to PayPal’s systems’ after discovery.

In a statement sent following publication, a PayPal spokesperson said ‘When there is a potential exposure of customer information, PayPal is required to notify affected customers. In this case, PayPal’s systems were not compromised. As such, we contacted the approximately 100 customers who were potentially impacted to provide awareness on this matter.’

Data potentially accessed includes names, email addresses, phone numbers, business addresses, Social Security numbers, and dates of birth. PayPal confirmed a small number of unauthorised transactions and said refunds were issued. Affected users had passwords reset and were offered credit monitoring.

Previous incidents include a 2023 credential stuffing attack that affected nearly 35,000 accounts and phishing campaigns that abused legitimate infrastructure. The company said it continues to use manual investigations and automated tools to mitigate fraud.

Customers are advised to use unique passwords, avoid unsolicited links, verify urgent messages directly via their accounts, and enable passkeys where available. Even limited breaches can heighten risks of targeted phishing and identity theft, especially for small businesses.

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.