Digital standards

Meta’s Behemoth AI model faces setback

Meta Platforms has postponed the release of its flagship AI model, known as ‘Behemoth,’ due to internal concerns about its performance, according to a report by the Wall Street Journal.

Instead of launching as planned, engineers are struggling to deliver improvements that would meaningfully advance the model beyond earlier versions.

Behemoth was originally scheduled for release in April to coincide with Meta’s first AI developer conference but was quietly delayed to June. The latest update suggests the launch has now been pushed to autumn or later, as internal doubts grow over whether it is ready for public deployment.

In April, Meta previewed Behemoth under the Llama 4 line, calling it ‘one of the smartest LLMs in the world’ and positioning it as a teaching model for future AI systems. Instead of Behemoth, Meta released Llama 4 Scout and Llama 4 Maverick as the latest iterations in its AI portfolio.

The delay comes amid intense competition in the generative AI space, where rivals like Google, OpenAI, and Anthropic continue advancing their models. Meta appears to be opting for caution instead of rushing an underwhelming product to market.

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

OpenAI launches AI safety hub

OpenAI has launched a public online hub to share internal safety evaluations of its AI models, aiming to increase transparency around harmful content, jailbreaks, and hallucination risks. The hub will be updated after major model changes, allowing the public to track progress in safety and reliability over time.

The move follows growing criticism about the company’s testing methods, especially after inappropriate ChatGPT responses surfaced in late 2023. Instead of waiting for backlash, OpenAI is now introducing an optional alpha testing phase, letting users provide feedback before wider model releases.

The hub also marks a departure from the company’s earlier stance on secrecy. In 2019, OpenAI withheld GPT-2 over misuse concerns. Since then, it has shifted towards transparency by forming safety-focused teams and responding to calls for open safety metrics.

OpenAI’s approach appears timely, as several countries are building AI Safety Institutes to evaluate models before launch. Instead of relying on private sector efforts alone, the global landscape now reflects a multi-stakeholder push to create stronger safety standards and governance for advanced AI.

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

TikTok adds AI tool to animate photos with realistic effects

TikTok has launched a new feature called AI Alive, allowing users to turn still images into dynamic, short videos. Instead of needing advanced editing skills, creators can now use AI to generate movement and effects with a few taps.

By accessing the Story Camera and selecting a static photo, users can simply type how they want the image to change — such as making the subject smile, dance, or tilt forward. AI Alive then animates the photo, using creative effects to produce a more engaging story.

TikTok says its moderation systems review the original image, the AI prompt, and the final video before it’s shown to the user. A second check occurs before a post is shared publicly, and every video made with AI Alive will include an ‘AI-generated’ label and C2PA metadata to ensure transparency.

The feature stands out as one of the first built-in AI image-to-video tools on a major platform. Snapchat and Instagram already offer AI image generation from text, and Snapchat is reportedly developing a similar image-to-video feature.

Meanwhile, TikTok is also said to be working on adding support for sending photos and voice messages via direct message — something rival apps have long supported.

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

NatWest hit by 100 million cyber attacks every month

NatWest is defending itself against an average of 100 million cyber attacks each month, according to the bank’s head of cybersecurity.

Speaking to Holyrood’s Criminal Justice Committee, Chris Ulliott outlined the ‘staggering’ scale of digital threats targeting the bank’s systems. Around a third of all incoming emails are blocked before reaching staff, as they are suspected to be the start of an attack.

Instead of relying on basic filters, NatWest analyses every email for malicious content and has a cybersecurity team of hundreds, supported by a multi-million-pound budget.

Mr Ulliott also warned of the growing use of AI by cyber criminals to make scams more convincing—such as altering their appearance during video calls to build trust with victims.

Police Scotland reported that cybercrime has more than doubled since 2020, with incidents rising from 7,710 to 18,280 in 2024. Officials highlighted the threat posed by groups like Scattered Spider, believed to consist of young hackers sharing techniques online.

MSP Rona Mackay called the figures ‘absolutely staggering,’ while Ben Macpherson said he had even been impersonated by fraudsters.

Law enforcement agencies, including the FBI, are now working together to tackle online crime. Meanwhile, Age Scotland warned that many older people lack confidence online, making them especially vulnerable to scams that can lead to financial ruin and emotional distress.

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

AI hallucination at center of Anthropic copyright lawsuit

Anthropic, the AI company behind the Claude chatbot, has been ordered by a federal judge to respond to allegations that it submitted fabricated material—possibly generated by AI—as part of its defense in an ongoing copyright lawsuit.

The lawsuit, filed in October 2023 by music publishers Universal Music Group, Concord, and ABKCO, accuses Anthropic of unlawfully using lyrics from over 500 songs to train its chatbot. The publishers argue that Claude can produce copyrighted material when prompted, such as lyrics from Don McLean’s American Pie.

During a court hearing on Tuesday in California, the publishers’ attorney claimed that an Anthropic data scientist cited a nonexistent academic article from The American Statistician journal to support the argument that Claude rarely outputs copyrighted lyrics.

One of the article’s alleged authors later confirmed the paper was a ‘complete fabrication.’ The judge is now requiring Anthropic to formally address the incident in court.

The company, founded in 2021, is backed by major investors including Amazon, Google, and Sam Bankman-Fried, the disgraced crypto executive convicted of fraud in 2023.

The case marks a significant test of how AI companies handle copyrighted content, and how courts respond when AI-generated material is used in legal proceedings.

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

Seattle startup ElastixAI raises $16 million for AI inference tech

A stealthy new AI startup in Seattle, ElastixAI, has raised $16 million to build technology that aims to reduce the cost and complexity of running large language models.

Rather than focusing on training, the company is developing an AI inference platform to optimise how these models operate, whether on cloud servers or edge devices. The funding round is led by Bellevue-based venture capital firm FUSE, with support from several others.

ElastixAI is led by CEO Mohammad Rastegari, formerly CTO of Xnor, a startup acquired by Apple in 2020. He co-founded the company with Saman Naderiparizi, also ex-Apple and Xnor, and Mahyar Najibi, who worked at both Apple and Waymo.

The team’s background in AI hardware and software gives them a unique edge in addressing challenges at a stage where AI models generate responses from trained data.

Instead of building a one-size-fits-all solution, the startup’s platform is designed for flexibility, allowing customers to fine-tune infrastructure to specific needs. ‘We saw a gap in delivering scalable and low-cost inference,’ said Rastegari.

The company remains in stealth but says its platform could serve both hyperscalers and enterprises looking to integrate AI into everyday operations.

With other players like Nvidia and Fireworks.ai competing in the inference space, ElastixAI may even count some of them as future customers.

Rastegari and Naderiparizi are also affiliate assistant professors at the University of Washington, and their startup reflects Seattle’s growing reputation as a hub for advanced AI development — a trend Apple has helped shape with several acquisitions in the region.

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

Valve denies Steam data breach

Valve has confirmed that a cache of old Steam two-factor authentication codes and phone numbers, recently circulated by a hacker known as ‘Machine1337’, is indeed real, but insists it did not suffer a data breach.

Instead of pointing to its own systems, Valve explained that the leak involves outdated SMS messages, which are typically sent unencrypted and routed through multiple providers. These codes, once valid for only 15 minutes, were not linked to specific Steam accounts, passwords, or payment information.

The leaked data sparked early speculation that third-party messaging provider Twilio was the source of the breach, especially after their name appeared in the dataset. However, both Valve and Twilio denied any direct involvement, with Valve stating it does not even use Twilio’s services.

The true origin of the breach remains uncertain, and Valve acknowledged that tracing it may be difficult, as SMS messages often pass through several intermediaries before reaching users.

While the leaked information may not immediately endanger Steam accounts, Valve advised users to remain cautious. Phone numbers, when combined with other data, could still be used for phishing attacks.

Instead of relying on SMS for security, users are encouraged to activate the Steam Mobile Authenticator, which offers a more secure alternative for account verification.

Despite the uncertainty surrounding the source of the breach, Valve reassured users there’s no need to change passwords or phone numbers. Still, it urged vigilance, recommending that users routinely review their security settings and remain wary of any unsolicited account notifications.

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

DeepMind unveils AlphaEvolve for scientific breakthroughs

Google DeepMind has unveiled AlphaEvolve, a new AI system designed to help solve complex scientific and mathematical problems by improving how algorithms are developed.

Rather than acting like a standard chatbot, AlphaEvolve blends large language models from the Gemini family with an evolutionary approach, enabling it to generate, assess, and refine multiple solutions at once.

Instead of relying on a single output, AlphaEvolve allows researchers to submit a problem and potential directions. The system then uses both Gemini Flash and Gemini Pro to create various solutions, which are automatically evaluated.

The best results are selected and enhanced through an iterative process, improving accuracy and reducing hallucinations—a common issue with AI-generated content.

Unlike earlier DeepMind tools such as AlphaFold, which focused on narrow domains, AlphaEvolve is a general-purpose AI for coding and algorithmic tasks.

It has already shown its value by optimising Google’s own Borg data centre management system, delivering a 0.7% efficiency gain—significant given Google’s global scale.

The AI also devised a new method for multiplying complex matrices, outperforming a decades-old technique and even beating DeepMind’s specialised AlphaTensor model.

AlphaEvolve has also contributed to improvements in Google’s hardware design by optimising Verilog code for upcoming Tensor chips.

Though not publicly available yet due to its complexity, AlphaEvolve’s evaluation-based framework could eventually be adapted for smaller AI tools used by researchers elsewhere.

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

Apple opens NFC chip, PayPal taps into it

PayPal has launched its tap-to-pay feature for iPhone users in Germany, allowing contactless payments at terminals that support Mastercard.

The rollout follows pressure from the European Union under the Digital Markets Act, which forced Apple to open up its NFC chip to third-party apps.

Currently, the feature is exclusive to iPhones and does not support Apple Watch. It mirrors earlier moves by other apps like Norway’s Vipps, which began using Apple’s newly accessible NFC hardware late last year.

Apple has also started expanding NFC access globally, enabling developers to integrate in-app payments and allowing businesses to accept contactless payments via iPhones using third-party apps like Venmo and PayPal Zettle.

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

Android adds new scam protection for phone calls

Google is introducing new protections on Android devices to combat phone call scams, particularly those involving screen-sharing and app installations. Users will see warning messages if they attempt to change settings during a call and Android will also block the deactivation of Play Protect features.

The system will now block users from sideloading apps or granting accessibility permissions while on a call with unknown contacts.

The new tools are available on devices running Android 16 and select protections are also rolling out to older versions, starting with Android 11

A separate pilot in the UK will alert users trying to open banking apps during a screen-sharing call, prompting them to end the call or wait before proceeding.

These features expand Android’s broader efforts to prevent fraud, which already include AI-based scam detection for phone calls and messages.

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