Access

Hackers use AI to create phishing sites in seconds

Hackers are now using generative AI tools to build convincing phishing websites in under a minute, researchers at Okta have warned. The company discovered that a tool developed by Vercel had been abused to replicate login portals for platforms such as Okta, Microsoft 365 and crypto services.

Using simple prompts like ‘build a copy of the website login.okta.com’, attackers can create fake login pages with little effort or technical skill. Okta’s investigation found no evidence of successful breaches, but noted that threat actors repeatedly used v0 to target new platforms.

Vercel has since removed the fraudulent sites and is working with Okta to create a system for reporting abuse. Security experts are concerned the speed and accessibility of generative AI tools could accelerate low-effort cybercrime on a massive scale.

Researchers also found cloned versions of the v0 tool on GitHub, which may allow continued abuse even if access to the original is restricted. Okta urges organisations to adopt passwordless systems, as traditional phishing detection methods are becoming obsolete.

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

AI model predicts sudden cardiac death more accurately

A new AI tool developed by researchers at Johns Hopkins University has shown promise in predicting sudden cardiac death among people with hypertrophic cardiomyopathy (HCM), outperforming existing clinical tools.

The model, known as MAARS (Multimodal AI for ventricular Arrhythmia Risk Stratification), uses a combination of medical records, cardiac MRI scans, and imaging reports to assess individual patient risk more accurately.

In early trials, MAARS achieved an AUC (area under the curve) score of 0.89 internally and 0.81 in external validation — both significantly higher than traditional risk calculators recommended by American and European guidelines.

The improvement is attributed to its ability to interpret raw cardiac MRI data, particularly scans enhanced with gadolinium, which are often overlooked in standard assessments.

While the tool has the potential to personalise care and reduce unnecessary defibrillator implants, researchers caution that the study was limited to small cohorts from Johns Hopkins and North Carolina’s Sanger Heart & Vascular Institute.

They also acknowledged that MAARS’s reliance on large and complex datasets may pose challenges for widespread clinical use.

Nevertheless, the research team believes MAARS could mark a shift in managing HCM, the most common inherited heart condition.

By identifying hidden patterns in imaging and medical histories, the AI model may protect patients more effectively, especially younger individuals who remain at risk yet receive no benefit from current interventions.

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

TikTok struggles to stop the spread of hateful AI videos

Google’s Veo 3 video generator has enabled a new wave of racist AI content to spread across TikTok, despite both platforms having strict policies banning hate speech.

According to MediaMatters, several TikTok accounts have shared AI-generated videos promoting antisemitic and anti-Black stereotypes, many of which still circulated widely before being removed.

These short, highly realistic videos often included offensive depictions, and the visible ‘Veo’ watermark confirmed their origin from Google’s model.

While both TikTok and Google officially prohibit the creation and distribution of hateful material, enforcement has been patchy. TikTok claims to use both automated systems and human moderators, yet the overwhelming volume of uploads appears to have delayed action.

Although TikTok says it banned over half the accounts before MediaMatters’ findings were published, harmful videos still managed to reach large audiences.

Google also maintains a Prohibited Use Policy banning hate-driven content. However, Veo 3’s advanced realism and difficulty detecting coded prompts make it easier for users to bypass safeguards.

Testing by reporters suggests the model is more permissive than previous iterations, raising concerns about its ability to filter out offensive material before it is created.

With Google planning to integrate Veo 3 into YouTube Shorts, concerns are rising that harmful content may soon flood other platforms. TikTok and Google appear to lack the enforcement capacity to keep pace with the abuse of generative AI.

Despite strict rules on paper, both companies are struggling to prevent their technology from fuelling racist narratives at scale.

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

Meta pursues two AI paths with internal tension

Meta’s AI strategy is facing internal friction, with CEO Mark Zuckerberg and Chief AI Scientist Yann LeCun taking sharply different paths toward the company’s future.

While Zuckerberg is doubling down on superintelligence, even launching a new division called Meta Superintelligence Labs, LeCun argues that even ‘cat-level’ intelligence remains a distant goal.

The new lab, led by Scale AI founder Alexandr Wang, marks Zuckerberg’s ambition to accelerate progress in large language models — a move triggered by disappointment in Meta’s recent Llama performance.

Reports suggest the models were tested with customised benchmarks to appear more capable than they were. That prompted frustration at the top, especially after Chinese firm DeepSeek built more advanced tools using Meta’s open-source Llama.

LeCun’s long-standing advocacy for open-source AI now appears at odds with the company’s shifting priorities. While he promotes openness for diversity and democratic access, Zuckerberg’s recent memo did not mention open-source principles.

Internally, executives have even discussed backing away from Llama and turning to closed models like those from OpenAI or Anthropic instead.

Meta is pursuing both visions — supporting LeCun’s research arm, FAIR, and investing in a new, more centralised superintelligence effort. The company has offered massive compensation packages to OpenAI researchers, with some reportedly offered up to $100 million.

Whether Meta continues balancing both philosophies or chooses one outright could determine the direction of its AI legacy.

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

DeepSeek gains business traction despite security risks

Chinese AI company DeepSeek is gaining traction in global markets despite growing concerns about national security.

While government bans remain in place across several countries, businesses are turning to DeepSeek’s models for low cost and firm performance, often ranking just behind OpenAI’s ChatGPT and Google’s Gemini in traffic and market share.

DeepSeek’s appeal lies in its efficiency. With advanced engineering techniques like its ‘mixture-of-experts’ system, the company has reduced computing costs by activating fewer parameters without a noticeable drop in performance.

Training costs have reportedly been as low as $5.6 million — a fraction of what rivals like Anthropic spend. As a result, DeepSeek’s models are now available across major platforms, including AWS, Azure, Google Cloud, and even open-source repositories like GitHub and Hugging Face.

However, the way DeepSeek is accessed matters. While companies can safely self-host the models in private environments, using the mobile app or website means sending data to Chinese servers, a key reason for widespread bans on public-sector use.

Individual consumers often lack the technical control enterprises enjoy, making their data more vulnerable to foreign access.

Despite the political tension, demand continues to grow. US firms are exploring DeepSeek as a cost-saving alternative, and its models are being deployed in industries from telecoms to finance.

Even Perplexity, an American AI firm, has used DeepSeek R1 to power a research tool hosted entirely on Western servers. DeepSeek’s open-source edge and rapid technical progress are helping it close the gap with much larger AI competitors — quietly but significantly.

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

Meta’s AI chatbots are designed to initiate conversations and enhance user engagement

Meta is training AI-powered chatbots that can remember previous conversations, send personalised follow-up messages, and actively re-engage users without needing a prompt.

Internal documents show that the company aims to keep users interacting longer across platforms like Instagram and Facebook by making bots more proactive and human-like.

Under the project code-named ‘Omni’, contractors from the firm Alignerr are helping train these AI agents using detailed personality profiles and memory-based conversations.

These bots are developed through Meta’s AI Studio — a no-code platform launched in 2024 that lets users build customised digital personas, from chefs and designers to fictional characters. Only after a user initiates a conversation can a bot send one follow-up, and that too within a 14-day window.

Bots must match their assigned personality and reference earlier interactions, offering relevant and light-hearted responses while avoiding emotionally charged or sensitive topics unless the user brings them up. Meta says the feature is being tested and rolled out gradually.

The company hopes it will not only improve user retention but also serve as a response to what CEO Mark Zuckerberg calls the ‘loneliness epidemic’.

With revenue from generative AI tools projected to reach up to $3 billion in 2025, Meta’s focus on more prolonged and engaging chatbot interactions appears to be as strategic as social.

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

X to test AI-generated Community Notes

X, the social platform formerly known as Twitter, is preparing to test a new feature allowing AI chatbots to generate Community Notes.

These notes, a user-driven fact-checking system expanded under Elon Musk, are meant to provide context on misleading or ambiguous posts, such as AI-generated videos or political claims.

The pilot will enable AI systems like Grok or third-party large language models to submit notes via API. Each AI-generated comment will be treated the same as a human-written one, undergoing the same vetting process to ensure reliability.

However, concerns remain about AI’s tendency to hallucinate, where it may generate inaccurate or fabricated information instead of grounded fact-checks.

A recent research paper by the X Community Notes team suggests that AI and humans should collaborate, with people offering reinforcement learning feedback and acting as the final layer of review. The aim is to help users think more critically, not replace human judgment with machine output.

Still, risks persist. Over-reliance on AI, particularly models prone to excessive helpfulness rather than accuracy, could lead to incorrect notes slipping through.

There are also fears that human raters could become overwhelmed by a flood of AI submissions, reducing the overall quality of the system. X intends to trial the system over the coming weeks before any wider rollout.

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

Croatia turns bus stops into AI-powered health stations

Croatia has introduced a groundbreaking public health campaign that uses AI and interactive billboards to tackle undiagnosed spinal issues. The initiative, led by Croatia Poliklinika, has provided over 100,000 free posture screenings and reached nearly 78% of the country’s population.

By converting bus stops into makeshift health stations, the project enables passers-by to receive posture checks via AI-powered digital billboards. The system analyses spine alignment at eight points and offers instant feedback, encouraging users to seek specialist help through a QR code.

The campaign has sparked immediate behavioural change, with 97% of people adjusting their posture during the scan. Since its launch, Croatia has seen a 46% rise in preventive spine check-ups, turning passive advertising spaces into tools for public well-being.

Croatia Poliklinika, which operates nine clinics nationwide, says the project demonstrates how digital infrastructure can reshape healthcare accessibility. The success is being hailed as a model for how technology can redefine public health engagement in everyday spaces.

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

Amazon reaches one million warehouse robots

Amazon has reached a major milestone with over one million robots now operating in its warehouses.

The one millionth robot, recently deployed to a facility in Japan, marks 13 years since the tech giant began introducing automation through its acquisition of Kiva Systems in 2012.

The robotic presence is fast approaching parity with Amazon’s human workforce, according to The Wall Street Journal. Robots now assist in around 75% of the company’s global deliveries.

The company continues to upgrade its robotic fleet, recently unveiling Vulcan — a dual-armed model equipped with a suction grip and a sense of touch to handle items more delicately.

Amazon is also introducing DeepFleet, a new generative AI model built using Amazon SageMaker.

Designed to optimise robotic movement within fulfilment centres, DeepFleet is expected to improve fleet speed by 10%. The model is trained on Amazon’s operational data, making it highly tailored to the company’s logistical network.

The expansion comes as Amazon opens next-generation fulfilment centres featuring ten times more robots instead of relying solely on existing warehouse models. The first of these facilities opened in late 2024 in Shreveport, Louisiana, signalling a shift toward even greater automation.

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

Grammarly invests in email with Superhuman acquisition

Grammarly announced on Tuesday that it has acquired email client Superhuman to expand its AI capabilities within its productivity suite.

Financial details of the deal were not disclosed by either company. Superhuman, founded by Rahul Vohra, Vivek Sodera and Conrad Irwin, has raised over $114 million from investors such as a16z and Tiger Global, with a last valuation of $825 million.

Grammarly CEO Shishir Mehrotra said the acquisition will enable the company to bring enhanced AI collaboration to millions more professionals, adding that email is not just another app but a crucial platform where users spend significant time.

Superhuman’s CEO Rahul Vohra and his team are joining Grammarly, promising to invest further in improving the Superhuman experience and building AI agents that collaborate across everyday communication tools.

Recently, Superhuman introduced AI-powered features like scheduling, replies and email categorisation. Grammarly aims to leverage the technology to build smarter AI agents for email, which remains a top use case for its customers.

The move follows Grammarly’s acquisition of productivity software Coda last year and the promotion of Shishir Mehrotra to CEO.

In May, Grammarly secured $1 billion from General Catalyst through a non-dilutive investment, repaid by a capped percentage of revenue generated using the funds instead of equity.

The Superhuman deal further signals Grammarly’s commitment to integrating AI deeply into professional communication.

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