Everything

Italian prosecutors confirm journalist was hacked with Paragon spyware

Italian authorities confirmed that a journalist who was alerted by WhatsApp last year of a suspected spyware attack on his phone was indeed hacked. 

In a press release sent to journalists on Thursday, the public prosecutors’ offices in Rome and Naples, which are investigating the spyware scandal in the country, said that a technical report concluded that the phones of journalist Francesco Cancellato, as well as Giuseppe Caccia and Luca Casarini, two immigration activists, all showed traces of having been infected with spyware on the “early hours” of December 14, 2024. 

“The execution of three consecutive attacks on the same night suggests that they may have been part of the same infection campaign,” the technical report said, according to the press release. 

The full report is not yet public.

This is the first independent confirmation that Cancellato, who is the director of the news website Fanpage, was hacked with spyware. In January 2025, Cancellato and around 90 other people, including journalists and members of civil society, were alerted by WhatsApp that they had been targeted with spyware made by Paragon Solutions, an Israeli-based company now owned by American private equity firm AE Industrial. 

According to the press release, Italian judicial authorities inspected the Paragon spyware server used by the intelligence agency, AISI, to target the phones of its targets. While the judicial authorities found evidence of operations against Caccia and Casarini, it found no evidence of an operation against Cancellato. 

It remains unclear who hacked Cancellato’s phone.

Contact Us
Do you have more information about Paragon, and this or other spyware campaigns? From a non-work device, you can contact Lorenzo Franceschi-Bicchierai securely on Signal at +1 917 257 1382, or via Telegram and Keybase @lorenzofb, or email.

By June 2025, an investigation by the Italian Parliamentary Committee for the Security of the Republic, known as COPASIR, concluded that Italian intelligence agencies had lawfully targeted Caccia and Casarini, but the committee found no evidence of a hack against Cancellato. 

The prosecutors’ offices said they will continue to investigate to identify Cancellato’s hackers. 

The Italian government, led by far-right prime minister Giorgia Meloni, has denied being behind the hack on Cancellato. In response to a question by the journalist during a press conference in January, Meloni only said that her government “is offering all its assistance and all the answers it can provide to help clarify this issue.”

The Italian government did not respond to TechCrunch’s request for comment.

“We are asking for clarity,” Cancellato said in an article on Thursday. “And we have not received it from the government, which has remained silent whenever possible for a year. And when it didn’t remain silent, it told lies.”

John Scott-Railton, one of the Citizen Lab researchers who investigated the Paragon cases in Italy, said that the new revelation about Cancellato’s hack “raises serious questions about why no confirmation was surfaced in prior official investigations by the Italian authorities.”

In response to the scandal, Paragon, whose spyware is called Graphite, cancelled its contracts with its Italian government customers.  

Spyware scandals spread across Europe

Apart from Caccia, Casarini, and Cancellato, there were several other people in Italy who were identified as spyware targets, including Ciro Pellegrino, who also works at Fanpage and was alerted of a suspected attack on his iPhone by Apple last year. Researchers at the Citizen Lab later concluded that Pellegrino was hacked with Paragon spyware.

The technical report mentioned by the prosecutor’s offices, however, said it only found evidence of spyware on the phones of Caccia, Casarini, and Cancellato, but not Pellegrino and another four people who are alleged victims. 

“I’m pretty disconcerted,” Pellegrino, who said he has not seen the full technical report yet, told TechCrunch. “How is it possible that Citizen Lab, an authority on spyware, found evidence that Paragon’s Graphite was on my phone, while the Italian prosecutors’ experts did not? And why would Apple send me the alerts? For fun?”

The prosecutor’s offices in Rome and Naples did not respond to a request for comment. 

A spokesperson for the Polizia Postale, which is investigating the case, referred TechCrunch to the prosecutor’s offices. 

Paragon, which as of last year had an active contract with the U.S. Immigration and Customs Enforcement (ICE), and REDLattice, a company that merged with the spyware maker after the acquisition by AE Industrial, did not respond to a request for comment. 

Italy is the most recent European country in recent years to have been embroiled in a spyware scandal, after similar cases in Greece, Hungary, Poland, and Spain. 

At the end of last month, a Greek court sentenced Tal Dilian and three other executives of the spyware maker Intellexa to eight years in prison for illegal wiretapping and privacy violations. 

The sentencing was part of the so-called “Greek Watergate” scandal, in which the Greek government was accused in 2022 of hacking the phones of politicians, journalists, businesspeople, and military officials with Intellexa’s spyware Predator. 

Read More

It’s official: The Pentagon has labeled Anthropic a supply-chain risk

The Department of Defense (DOD) has officially notified Anthropic leadership that the company and its products have been designated a supply-chain risk, Bloomberg reports, citing a senior department official. 

The designation comes after weeks of conflict between the AI lab and the DOD. Anthropic CEO Dario Amodei has refused to allow the military to use its AI systems for mass surveillance of Americans or to power fully autonomous weapons with no humans assisting in the targeting or firing decisions. The Department has argued that its use of AI should not be limited by a private contractor. 

Supply-chain-risk designations are typically reserved for foreign adversaries. The label requires any company or agency that does work with the Pentagon to certify that it doesn’t use Anthropic’s models. 

The Pentagon’s finding threatens to disrupt both the company and its own operations. Anthropic has been the only frontier AI lab with classified-ready systems. The U.S. military is currently relying on Claude in its Iran campaign, where American forces are using AI tools to quickly manage the data for their operations. Claude is one of the main tools installed in Palantir’s Maven Smart System, which military operators in the Middle East rely on, according to Bloomberg.

Labeling Anthropic a supply-chain risk over this disagreement is an unprecedented move from the Department, several critics say. Dean Ball, a former Trump White House AI adviser, has referred to the designation as a “death rattle” of the American republic, arguing government has abandoned strategic clarity and respect in favor of “thuggish” tribalism that treats domestic innovators worse than foreign adversaries. 

Hundreds of employees from OpenAI and Google have urged the DOD to withdraw its designation and called on Congress to push back on what could be perceived as an inappropriate use of authority against an American technology company. They have also urged their leaders to stand together to continue to refuse the DOD’s demands to use their AI models for domestic mass surveillance and “autonomously killing people without human oversight.”

TechCrunch has reached out to Anthropic for comment.

Techcrunch event

San Francisco, CA
|
October 13-15, 2026

In the midst of the dispute, OpenAI forged its own deal with the Department to allow the military to use its AI systems for “all lawful purposes.” Some of the company’s employees have expressed concern about the ambiguous phrasing of the deal, which could lead to exactly the type of uses Anthropic was trying to avoid.

Amodei has called the actions of the DOD “retaliatory and punitive,” and reportedly said his refusal to praise or donate to President Trump contributed to the dispute with the Pentagon. OpenAI president Greg Brockman has been a staunch backer of Trump, recently donating $25 million to the MAGA Inc. Super PAC. 

Read More

X revamps Creator Subscriptions with new features, like exclusive threads and shareable cards

Elon Musk-owned X announced on Thursday that it’s revamping the social network’s Creator Subscriptions offering with a number of new features, including exclusive threads, a refreshed subscriptions paywall, a new dashboard, and a shareable subscriptions card, among other changes.

The company believes these updates will allow creators to attract more fans and provide better experiences, as it pushes to become a bigger part of the creator economy alongside Instagram, TikTok, YouTube, and Facebook.

X also highlighted the recent launch of its “Paid Partnership” label, which is applied to posts where the content is paid for or incentivized by a third party.

One key new feature arriving Thursday is the launch of “exclusive threads,” which allow creators to monetize their posts directly on X, instead of sending fans to an external website or newsletter. With these posts, creators can tease the value of the thread in a parent post, then allow audiences to subscribe directly to the individual conversation. Subscriber-only content will now live in the account’s main profile feed, instead of in a separate Subscriptions tab, according to X.

Other new tools are designed to help creators market themselves and track their analytics. This includes a new Subscriptions card that can be shared on X to promote the creator’s content. There’s also a new dashboard that lets creators more easily track their earnings alongside other subscriber insights, as well as access built-in growth tools.

Image Credits:TechCrunch (screenshot)

X is also tweaking the subscription process, which includes a refreshed paywall where creators can highlight the benefits of their subscriptions and a faster onboarding process for setting up subscriptions that now takes just two steps. X said it has also sped up application review times, so creators can more quickly get started with monetization.

“We are doubling down on creators, the platform’s most influential experts, thought leaders, and voices, by investing in tools that enable sustainable incomes and deeper connections with their audiences,” said Allegra Jacchia, Creators Product at X, in a statement about the launch. “We believe creators deserve to be compensated for the value, knowledge, and perspectives they bring to the platform, and a big part of that is enabling them to monetize their most loyal supporters directly.”

The company also formally announced the recent launch of the “Paid partnership” label, which offers creators a built-in tool to comply with X’s policy and industry regulations over labeling ads, without having to use hashtags. Instead, the new label appears directly below the post’s content, and it can be added after the fact if a creator forgets to include it when they first post.

Image Credits:TechCrunch (screenshot)

X’s Creator Revenue Sharing program, which launched globally in 2023, allows creators to generate income for their popular posts. While designed to boost engaging content, some have criticized the program’s design, saying it incentivizes creators to post sensationalized content, including clickbait or ragebait. However, X has noted that some of its top creators include stock traders offering analysis, like @Banana3Stocks and @Jake_Wujastyk.

The company has paid out more than $45 million to creators to date, according to a post from Grok, the AI assistant integrated into X. For 2026, the company has more than doubled the revenue pool available due to X’s Premium subscription growth, Grok said. A spokesperson for X confirmed the latter, pointing to an X Creators post that also referenced changes to how payouts were calculated.

In January, some creators posted screenshots of their earnings, suggesting that payouts had recently increased. X also recently pushed creators to try long-form articles with a recent $1 million campaign for the top article.

The changes to the subscriptions arrive amid a flurry of news from X this week, which also included the beta tests of X’s stand-alone app, XChat, and payments service X Money. In addition, the company announced it would suspend creators from its revenue-sharing program if they posted unlabeled AI videos depicting armed conflict.

X has long sought to attract creators to its platform, even before the company was acquired by Musk. However, the majority of X’s revenue today still comes from advertisers, not premium subscriptions or creator content.

Read More

Hardware testing startup Nominal hits $1B valuation, raises $155M in 10 months

Nominal on Thursday announced a fresh $80 million Series B extension round at a $1 billion valuation, led by Founders Fund. This followed the company’s $75 million Series B round led by Sequoia in September.

The company offers software that helps hardware engineers test their designs and began as a picks-and-shovels type of startup for the defense industry. The three-and-a-half-year-old, L.A.-based company says that in the last 10 months, it has landed four of the five largest defense contractors as customers.

CEO and co-founder Cameron McCord (pictured) is a former U.S. Navy submarine officer and an alum of defense tech company Anduril, which is also one of Nominal’s marquee customers. So perhaps it’s no surprise that Anduril co-founder and executive chairman Trae Stephens, who is also a Founders Fund partner, led this as a preemptive deal, McCord told Bloomberg and confirmed to TechCrunch. Sequoia, General Catalyst, Lux Capital, Red Glass, and Lightspeed also participated.

Next up, Nominal plans to expand beyond defense tech and into industrial sectors like automotive, robotics, and other industries. And it’s made a good start. The company tells TechCrunch that some of its other customers include Pratt Miller Motorsports (the Corvette Racing Team) and nuclear energy company Antares.

Read More

Roblox launches real-time AI chat rephrasing to filter out banned language

Roblox is introducing a real-time, AI-powered chat rephrasing feature that automatically replaces banned words with more respectful language, the company announced on Thursday.

The new feature goes beyond Roblox’s current text filter, which simply replaces banned words and phrases with the “#” symbol. But Roblox says that when users encounter strings of “####,” conversations can be disrupted and hard to follow.

Now, instead of simply displaying hash marks, filtered text will be rephrased into more respectful language that is close to the user’s original intent.

For example, a message that reads “Hurry TF up!” would previously have appeared as “####,” but will now be rephrased to “Hurry up!” Everyone in the chat is notified that the message has been rephrased to keep the conversation civil, Roblox says.

“Chat is central to how people connect, coordinate, and play on Roblox,” said Rajiv Bhatia, vice president of User and Discovery Product at Roblox, in a press release. “Real-time rephrasing helps keep gameplay and conversations on track while guiding language toward what’s appropriate. This approach reduces friction in chat while maintaining the standards that help keep our community civil.”

Roblox notes that while rephrasing reduces some of the disruption in the chat, its safety system remains in effect for more serious behavior. 

The new feature is supported in all languages currently available through Roblox’s automatic translation tools.

Techcrunch event

San Francisco, CA
|
October 13-15, 2026

The company is also upgrading the text-filtering system to detect more variations of banned language. Roblox says early results show that the system is getting better at detecting leetspeak, and more sophisticated attempts to bypass filters. This has allowed the company to reduce the prevalence of false negatives when sharing or soliciting personal information by 20x.

Thursday’s announcement comes shortly after Roblox introduced mandatory facial verification for access to chats on its platform following a wave of lawsuits over child safety from the attorneys general of Texas, Kentucky, and Louisiana, among others. The lawsuits were filed in response to reports that Roblox was exposing young users to dangerous risks, such as grooming and explicit content.

Read More

EXCLUSIVE: Luma launches creative AI agents powered by its new ‘Unified Intelligence’ models

AI video-generation startup Luma on Thursday launched Luma Agents, designed to handle end-to-end creative work across text, image, video, and audio. Luma Agents are powered by the startup’s Unified Intelligence family of models, with architecture trained on a single multimodal reasoning system. 

Luma Agents are being pitched as a new way of doing work for ad agencies, marketing teams, design studios, and enterprises. Luma says its agents are capable of planning and generating text, image, video, and audio while coordinating with other AI models, including Luma’s Ray 3.14, Google’s Veo 3 and Nano Banana Pro, ByteDance’s Seedream, and ElevenLabs’ voice models. 

Luma’s agents are built on the startup’s Uni-1 model, the first of its Unified Intelligence family of AI models. It has been trained on audio, video, image, language, and spatial reasoning, according to Amit Jain, chief executive officer and co-founder of Luma.

Jain told TechCrunch that the Uni-1 model can “think in language and imagine and render in pixels or images … we call it ‘intelligence in pixels.’” Other output capabilities like audio and video will come in subsequent model releases, he added. 

“Our customers aren’t buying the tool; they’re redoing how business is done,” Jain said. 

Image Credits:Luma AI

Luma has already started rolling out its new agentic platform with existing customers, including global ad agencies Publicis Groupe and Serviceplan, as well as for brands like Adidas, Mazda, and Saudi AI company Humain. 

Jain said the Luma Agents are a game changer because they can maintain persistent context across assets, collaborators, and creative iterations. They can also evaluate and refine outputs, improving their own results through an iterative self-critique, according to Jain. 

Techcrunch event

San Francisco, CA
|
October 13-15, 2026

This sort of check-your-work capability is what has made coding agents so useful, Jain said. “You need that ability to evaluate your work, fix it, and do that loop until the solution is good and accurate.”

Jain said the current workflow for using AI tools in creative environments doesn’t have the same acceleration of benefits people in the creative industry expect from AI. Instead, it’s more like: “Here are 100 models. Learn how to prompt them,” he said.

He said what makes Luma Agents different is that you don’t need to prompt back and forth for each iteration on an image or idea — the system instead generates large sets of variations and lets users steer the direction through conversation.

“With Unified Intelligence, because these models understand in addition to being able to generate, we are able to build a system that is able to do this sort of end-to-end work,” Jain said. 

Take, for instance, a human architect designing a building. As they draw the lines, they are creating an internal mental representation of the structure, light, spatial dynamics, and lived experience. This, Jain says, is the same principle upon which Unified Intelligence is built. 

Jain said the system could significantly speed up creative workflows. In a demonstration, he showed how a 200-word brief and an image of a product (a tube of lipstick) led the system to generate various ideas for locations, models, and color schemes for an ad campaign. 

In another example, Luma Agents turned a brand’s $15 million, year-long ad campaign into multiple localized ads for different countries in 40 hours for under $20,000, passing the brand’s internal quality controls and accuracy checks, Jain said. 

While Luma Agents is now publicly available via API, Jain said the startup plans to roll out access gradually to ensure users maintain reliable access and avoid workflow disruptions.

Read More

Google says half of all zero-days it tracked in 2025 targeted buggy enterprise tech

A new report by Google found that about half of the zero-day bugs it tracked last year exploited enterprise devices, marking a new high for hackers who are increasingly finding new ways to target large companies and steal their data.

According to the search and security giant’s annual report, 48% of the tracked zero-days — vulnerabilities in software that are unknown to its maker at the time they are exploited — were found in technologies used by corporations and large businesses. About half of those zero-days exploited the very devices that are designed to protect enterprise networks from digital intruders.

Google said security and networking devices, such as firewalls made by Cisco and Fortinet, and VPN and virtualization platforms like Ivanti and VMware, were among the top targeted vendors last year. All four of the companies said hackers have exploited their products on customer networks in recent months.

Google’s researchers said that hackers exploited common flaws, like input validation and incomplete authorization processes, to break through firewall and VPN defenses to gain access to customer networks. These classes of bugs are generally easier to exploit, but typically require a software update to fix. 

The company also pointed to other buggy software that makes up the remaining half of enterprise zero-days. Google noted the Clop extortion gang’s campaign against Oracle E-Business Suite customers, which allowed hackers to walk away with reams of human resources data from dozens of companies about their staff and executives. The hacks affected Harvard University, the American Airlines subsidiary Envoy, and The Washington Post, among others.

The remaining 52% of zero-day bugs were found in consumer and end-user products, such as those made by Microsoft, Google, and Apple, according to the report. Most of the zero-days in consumer software were found in operating systems, with mobile devices also seeing more zero-days than in previous years.

Google said it also attributed more zero-days to surveillance vendors than traditional government-backed espionage groups. Surveillance vendors are typically spyware makers and exploit developers, which work on behalf of governments to hack into people’s phones. Google said this shift demonstrated “a slow but sure movement in the landscape” in how governments seek access to hacking tools.

Techcrunch event

San Francisco, CA
|
October 13-15, 2026

Read More

OpenAI launches GPT-5.4 with Pro and Thinking versions

On Thursday, OpenAI released GPT-5.4, a new foundation model billed as “our most capable and efficient frontier model for professional work.” In addition to the standard version, GPT-5.4 is also available as a reasoning model (GPT-5.4 Thinking) or optimized for high performance (GPT-5.4 Pro).

The API version of the model will be available with context windows as large as 1 million tokens, by far the largest context window available from OpenAI.

OpenAI also emphasized improved token efficiency, saying GPT-5.4 was able to solve the same problems with significantly fewer tokens than its predecessor.

The new model comes with significantly improved benchmark results, including record scores in computer use benchmarks OSWorld-Verified and WebArena Verified. The new model also scored a record 83% on OpenAI’s GDPval test for knowledge work tasks.

GPT-5.4 also took the lead on Mercor’s APEX-Agents benchmark, designed to test professional skills in law and finance, according to a statement from Mercor CEO Brendan Foody.

“[GPT-5.4] excels at creating long-horizon deliverables such as slide decks, financial models, and legal analysis,” Foody said in the statement, “delivering top performance while running faster and at a lower cost than competitive frontier models.”

GPT-5.4 continues the company’s efforts to limit hallucinations and factual errors. OpenAI said the new model was 33% less likely to make errors in individual claims when compared to GPT 5.2, and overall responses were 18% less likely to contain errors.

Techcrunch event

San Francisco, CA
|
October 13-15, 2026

As part of the launch, OpenAI has reworked how the API version of GPT-5.4 manages tool calling, introducing a new system called Tool Search. Previously, system prompts would lay out definitions for all available tools when calling the model — a process that could consume a lot of tokens as the number of available tools grew. The new system allows models to look up tool definitions as needed, resulting in faster and cheaper requests in systems with many available tools.

OpenAI has also included a new safety evaluation to test its models’ chain-of-thought, the running commentary given by the models to show thought process through multi-step tasks. AI safety researchers have long worried that reasoning models could misrepresent their chain-of-thought, and testing shows it can happen under the right circumstances.

OpenAI’s new evaluation shows that deception is less likely to happen in the Thinking version of GPT-5.4, “suggesting that the model lacks the ability to hide its reasoning and that CoT monitoring remains an effective safety tool.”

Read More

Cursor is rolling out a new kind of agentic coding tool

As agentic coding spreads, the working life of a software engineer has become dazzlingly complex. A single engineer might oversee dozens of coding agents at once, launching and guiding different processes as necessary.

It’s a lot to keep track of, and human engineers’ attention has quickly become the limiting resource.

Cursor launched a new tool Thursday aimed at keeping that chaos in check. Called Automations, the new system gives users a way to automatically launch agents within their coding environment, triggered by a new addition to the codebase, a Slack message, or a simple timer. As Cursor describes it, it’s a way to review and maintain all the new code created by agentic tools — without tracking dozens of agents at once.

At the most basic level, Automations are a way for engineers to break out of the “prompt-and-monitor” dynamic that defines most agent-based engineering. Instead of launching agents with a human prompt, Cursor’s Automation framework lets you launch agents automatically — and loop humans in whenever they’re needed.

“It’s not that humans are completely out of the picture,” Jonas Nelle, Cursor’s engineering chief for asynchronous agents, told TechCrunch in an interview. ”It’s that they aren’t always initiating. They’re called in at the right points in this conveyor belt.”

One early example is Bugbot, a long-standing Cursor feature that the team sees as a predecessor to the broader Automation system. The Bugbot system is triggered every time an engineer makes an addition to the codebase and reviews the new code for bugs and other issues. Using Automations, Cursor has been able to expand that system to more involved security audits and more thorough reviews.

“This idea of thinking harder, spending more tokens to find harder issues, has been really valuable,” said  engineering lead Josh Ma.

Techcrunch event

San Francisco, CA
|
October 13-15, 2026

Cursor estimates that it runs hundreds of automations per hour, reaching far beyond simple code review. The system is also used for incident response, with PagerDuty incidents initiating an agent that can immediately query server logs through an MCP connection. A separate automation offers weekly summaries of changes to the codebase on Cursor’s company Slack.

“In the abstract, anything that an automation kicks off, a human could have also kicked off,” said Nelle. “But by making it automatic, you change the types of tasks that models can usefully do in a codebase.”

The new system comes amid intense competition in the agentic coding space, with both OpenAI and Anthropic having made significant updates to their agentic coding tools in the past month. 

Ramp data shows Cursor’s market share holding steady since May, with roughly 25% of generative AI clients subscribing to Cursor in some capacity.

Still, the overall growth of the agentic coding space has kept the company’s revenue increasing at a stunning pace. Earlier this week, Bloomberg reported that Cursor’s annual revenue had grown to more than $2 billion, doubling over the past three months.

Read More

Meta sued over AI smart glasses’ privacy concerns, after workers reviewed nudity, sex, and other footage

Meta is facing a new lawsuit over its AI smart glasses and their lack of privacy, after an investigation by Swedish newspapers found that workers at a Kenya-based subcontractor are reviewing footage from customers’ glasses, which included sensitive content, like nudity, people having sex, and using the toilet.

Meta claimed it was blurring faces in images, but sources disputed that this blurring consistently worked, reports noted. The news prompted the U.K. regulator, the Information Commissioner’s Office, to investigate the matter.

Now, the tech giant is facing a lawsuit in the United States, as well. In the newly filed complaint, plaintiffs Gina Bartone of New Jersey and Mateo Canu of California, represented by the public interest-focused Clarkson Law Firm, allege that Meta violated privacy laws and engaged in false advertising.

The complaint alleges that the Meta AI smart glasses are advertised using promises like “designed for privacy, controlled by you,” and “built for your privacy,” which might not lead customers to assume their glasses’ footage, including intimate moments, was being watched by overseas workers. The plaintiffs believed Meta’s marketing and said they saw no disclaimer or information that contradicted the advertised privacy protections.

The suit charges Meta and its glasses manufacturing partner Luxottica of America with conduct that violates consumer protection laws. Meta does not have a comment on the litigation at this time.

Clarkson Law Firm, which over the years has filed other major lawsuits against tech giants, including Apple, Google, and OpenAI, points to the scale of the issues at hand. In 2025, over seven million people bought Meta’s smart glasses, which means their footage is fed into a data pipeline for review, and they can’t opt out.

Meta told the BBC that when people share content with Meta AI, it uses contractors to review the information to improve people’s experience with the glasses, which is explained in its privacy policy, and pointed to Supplemental Meta Platforms Terms of Service, without specifying where this was noted. The news outlet, however, found that a mention of human review could be found in Meta’s U.K. AI terms of service.

Techcrunch event

San Francisco, CA
|
October 13-15, 2026

A version of that policy that applies to the U.S. states “In some cases, Meta will review your interactions with AIs, including the content of your conversations with or messages to AIs, and this review may be automated or manual (human).”

Read More