Malicious packages on the Node Package Manager (npm) and the Python Package Index (PyPI) delivered stealer malware to developers and users of Paysafe, Skrill, and Neteller payment applications.

The threat actor published at least 17 malicious packages simultaneously, each tasked to exfiltrate credentials and access tokens to a command-and-control server hosted on Amazon Web Services (AWS).

All three payment platforms are popular, with Paysafe being mostly used by e-commerce sites and online marketplaces, gaming platforms, travel businesses, and financial services or software-as-a-service (SaaS) providers.

Skrill and Neteller are digital wallets and money transfer services used in online betting, cryptocurrency exchanges, and on Forex trading platforms.

Software developers working on such platforms integrate Paysafe’s SDKs into apps and websites to implement a secure payments and funds management system.

According to application security company Socket, these developers are the targets of the latest campaign via the following packages:

1. npm/paysafe-checkout
2. npm/paysafe-vault
3. npm/neteller
4. npm/skrill-payments
5. npm/paysafe-js
6. npm/paysafe-api
7. npm/paysafe-node
8. npm/paysafe-cards
9. npm/paysafe-fraud
10. npm/paysafe-kyc
11. npm/skrill
12. npm/skrill-sdk
13. npm/paysafe-payments
14. pypi/paysafe-kyc
15. pypi/paysafe-payments
16. pypi/paysafe-sdk
17. pypi/paysafe-api

The researchers say that the 13 npm packages published four malicious versions, from 1.0.0 to 1.0.3, whereas the PyPI packages published only one malicious version, 1.0.0.

All 17 packages pretend to be legitimate payment SDKs, even exposing the expected APIs, but instead return fake success responses rather than communicate with Paysafe’s backend services.

The real purpose is credential theft, as the embedded malicious code searches compromised environments for secrets such as tokens, passwords, and API keys.

According to Socket, the exfiltrated data includes Paysafe API keys, AWS keys, GitHub tokens, npm tokens, hostname, username, and metadata about API usage.

The data theft module in the npm packages attempts exfiltration only if a Paysafe API key is present and activates when the fake SDK is called.

The PyPI packages automatically activate the data theft routine upon initialization and do not require a Paysafe API key to be present at all.

Socket’s analysis of the malware reveals that it includes some rather basic anti-analysis features, stopping execution if it detects fewer than 2 CPU cores or if the hostname or username contains cues indicating a virtualized environment.

It is unclear who is behind this campaign, but Socket’s report highlights some attributes suggesting that the threat actor is sufficiently technical and may return in a more organized way.

The researchers warn that the attacker’s ability to pivot between ecosystems may make it more difficult to defend if there is only one ecosystem of visibility.

If any of the listed packages were installed, developers are recommended to immediately “rotate all secrets on any machine that imported or executed this package.”

 The researchers also advise searching dependency trees for the package names used in the campaign and deny any requests for them at the registry proxy level.

It is also recommended to look in the logs of Continuous Integration (CI) systems for PAYSAFE_API_KEY in combination with any of the listed package names.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Get the whitepaper

A semi-trailer that helps propel itself entered commercial road testing in late May, when a powertrain kit developed by Nivalis Energy Europe, headquartered in Luxembourg with engineering operations in Germany, was fitted to a trailer supplied by Amsterdam-based TIP Group. The self-powered trailer was handed over to German transport operator Sommer for use in its working fleet.

The Nivalis Powered Trailer Kit centers on an electric axle co-developed with Wiehl, Germany–based running gear specialist BPW, rated at 50 kilowatts peak, capable of both propulsion assistance and regenerative braking. That axle draws on a 60-kilowatt-hour, 400-volt lithium-ion battery pack charged from three sources: the axle itself during braking and deceleration, a full-rooftop array of photovoltaic panels generating up to 3.7 kilowatts-peak, and a 32-amp, three-phase AC grid connection available during parking stops. The driver’s only window into the system is a small display readable from the cab’s side mirror that shows the system status and battery charge level. Nothing about the trailer’s handling or licensing requirements changes.

The partners project savings of up to 7,000 liters of diesel per trailer per year, which is enough to keep about 19 tonnes of carbon dioxide out of the air. These figures are based on a trailer running 100,000 kilometers annually at payloads between 20 and 24 tonnes, on a mix of long-haul and hub-to-hub routes.

Pavel Gilman, vice president of sales and marketing at Nivalis, breaks down where those savings come from: roughly 30 to 35 percent from the electric axle during braking and deceleration, 11 to 15 percent from the rooftop solar panels, and the remainder (roughly half) from grid charging during parking stops. The pilot is planned to run for more than a year, spanning multiple seasons. The retrofit cost has not been disclosed, and the pilot is running on a single trailer. But the underlying assumptions are now on the table and they represent a specific, high-utilization use case (meaning a truck that’s almost always on the move, filled to capacity with freight) not a universal one.

Across Europe and North America, a growing number of companies have concluded that electrifying the trailer, rather than replacing the tractor unit, may be the fastest and most cost-effective path to decarbonizing long-haul freight. A new battery-electric heavy truck carries a high upfront cost and demands charging infrastructure that most freight corridors do not yet reliably provide. A retrofit kit fitted to an existing trailer is meant to sidestep both problems.

The question the industry has been working to answer is whether the energy harvested from regenerative braking, rooftop solar, and grid charging in short bursts when the vehicle is parked for loading and unloading is enough to produce savings that recover the kit’s cost in a reasonable timeframe. Several companies now believe the answer is yes, and they are accumulating field data to prove it—though not all of them are going about it the same way.

Trailer industry places its bets

The competitive landscape has taken shape most visibly in Germany. Trailer Dynamics, an Aachen-based company, has conducted field tests with BMW Logistics, DB Schenker, Duvenbeck, and Volkswagen Konzernlogistik, reporting average fuel savings of around 40 percent for diesel tractor combinations, substantially higher than the up to 18 percent reduction implied by the Nivalis projection. The difference traces directly to battery size, but Trailer Dynamics frames the choice as an economic question rather than an architectural one.

“The discussion should not start with battery size, but with the economics of the transport operation,” the company said in response to written questions. “There is no single battery capacity that is universally right for every fleet.”

Trailer Dynamics’s modular system offers three configurations ranging from 187 to 551 kilowatt-hours, sized to match route profile, annual mileage, payload, and charging access. The M300 version, whose designation reflects the capacity of its 300-kilowatt-hour lithium iron phosphate battery supplied by Chinese battery manufacturer CATL, adds approximately four tonnes to the trailer, roughly three times the one-to-1.4–tonnes added to a trailer by the Nivalis system.

Both companies’ systems would extend the range of a battery-electric tractor by reducing the energy demand on the tractor’s motor. But Trailer Dynamics explicitly targets that use case, claiming its self-propelled trailer yields combined ranges of up to 850 kilometers—enough to eliminate intermediate charging stops on many long-haul routes. Nivalis has not published range extension figures for electric tractor combinations, and its smaller battery and peak lower output suggest the effect would be more modest.

That higher energy storage capability widens the addressable market for Trailer Dynamics considerably and helps explain the investment flowing into the self-propelled trailer space. In November 2025, the European Investment Bank extended a €25 million loan to the company, backed by the European Union’s InvestEU program, to support commercialization. Trailer Dynamics says it plans to begin industrial-scale production in 2028, with adoption expected to accelerate as European carbon dioxide reduction requirements tighten toward 2030.

ZF, the German automotive supplier, entered the space with its TrailTrax system, using an electric axle rated at up to 210 kilowatts continuous power. ZF claims that, between onboard battery storage and energy recovered via regenerative braking, the self-propelled trailer system yields up to 16 percent in energy and carbon dioxide savings when combined with an ICE powered truck. The company also says TrailTrax can reduce carbon dioxide emissions by as much as 40 percent with opportunistic plug-in charging. Trailer manufacturers Kässbohrer and Krone have adopted the platform, as has BPW—the same running gear specialist co-developing the Nivalis axle.

In North America, Range Energy is developing a system with up to 300 kilowatt-hours of onboard energy capacity, compatible with diesel, battery-electric, and hydrogen fuel cell tractors. Range, which has announced a partnership with ZF, to help drive the development and adoption of the Range eTrailer System within the North American commercial trucking industry, is now equipping its trailers with ZF’s AxTrax 2 e-axle for battery-powered propulsion. Range Energy has a separate pilot agreement with DB Schenker, the German logistics company that is also among the European operators that tested the Trailer Dynamics system. Range and DB Schenker say they plan to deploy a powered trailer in commercial trucking operations in North America, with first deliveries scheduled for later this year. The breadth of activity across continents reflects a field that has moved well past the question of whether powered trailers work. The argument now is about which architecture works best and at what cost.

What the field does not yet have is a common standard for measuring and reporting savings. The figures from various pilots—an average of 40 percent from Trailer Dynamics, up to 18 percent implied by the Nivalis projection—reflect different routes, loads, seasons, and battery sizes. In some cases, they represent short validation runs rather than sustained operational data. Fleet operators evaluating competing systems are working with numbers that are difficult to interpret and impossible to rank against one another.

Both architectures reduce available payload, but by very different margins. The M300’s roughly four-tonne addition dwarfs the one-to-1.4-tonne addition of the Nivalis system. Trailer Dynamics argues the weight penalty is largely academic in practice, because more than 90 percent of trailer movements are constrained by cargo volume before they approach legal weight limits. Under current European regulations, both systems reduce payload on a one-for-one basis. Frameworks under discussion would change that. New rules could allow up to four extra tonnes for electric trucks, with proposals to extend the provision to electric trailers. If amended, the payload effect would turn positive for both systems. Until then, every kilogram of kit is a kilogram unavailable for freight.

Small versus large battery systems

The choice between large-battery and small-battery powered trailers is a bet on which cost will fall faster: battery pack prices or the cost of grid charging infrastructure. A large-battery system delivers higher savings but requires reliable charging access across the operating cycle. If infrastructure buildout stalls—as it has repeatedly in heavy-duty transport—operators face the same dependency problem that has slowed battery-electric truck adoption. The Nivalis architecture hedges against that risk: its 32-amp connection requires only a standard industrial outlet, and the solar array and regenerative braking handle significant energy input without infrastructure at all. Gilman frames the design philosophy in terms of the industry it serves.

“Logistics lives with low margins,” he said. “We are focused on the product which fits the industry technically and financially. It overcomes the capital expenditure hurdle and maximizes financial benefit by adding sources of energy which are symbiotic to each other.” And because Nivalis’s axle is comparatively light, he says, operators won’t be forced to reduce payload.

Trailer Dynamics sees it differently.

“Long-haul transport will increasingly move toward depot-based and destination-based charging models,” says Michael W. Nimtsch, the company’s Managing Director. “The question is not how small a battery can be made, but how much economic value each additional kilowatt-hour can generate over the life of the vehicle.”

On solar and regenerative recovery, Nimtsch argues both are useful complements to stored battery energy rather than substitutes for it.

“Compared with the daily energy demand of a long-haul truck, solar generation remains relatively modest,” he says. The Nivalis energy breakdown supports that view in relative terms: Grid charging contributes the largest share of projected savings, regenerative braking second, and solar third. That hierarchy means performance depends more on charging access during dwell time than the multi-source framing might suggest even if that access requires only a standard industrial outlet.

Trailer Dynamics prices its system between €145,000 and €195,000 and targets a payback period of no more than five years. Nivalis targets five to six years at current costs, falling to three to four years as volumes grow. Asked exactly what the price tag says, the company declined to answer. The minimum annual savings needed, Gilman said, is between €5,000 and €6,000 per trailer. Until someone publishes a full year of results from a trailer running in normal commercial rotation, fleet operators cannot answer the two questions that actually drives adoption: What does this cost, and when does it pay back?

On Wednesday, the US Federal Trade Commission announced a settlement with tractor manufacturer John Deere over a 2025 lawsuit that accused the company of behavior that “unlawfully acquired and maintained monopoly power in markets for repair services for Deere farm equipment.”

The full statement lays out obligations for John Deere’s repair services, requiring the company to give farmers and third-party repair shops access to the same equipment and repair resources it provides to official John Deere dealers. This includes software capabilities, such as reading and resetting codes and pairing with other software, which customers have long had limited access to, creating delays when diagnosing equipment problems. Delayed fixes can mean delayed harvests, which many farmers saw as a fundamental threat to their livelihoods.

Under the agreement, John Deere will be required to provide this level of access, equipment, and services for the next 10 years, monitored by the FTC.

“After years of fighting for the right to repair, this order gives farmers real hope,” Willie Cade, a board member of the repair advocacy organization Repair.org, wrote in an email to WIRED. “But promises on paper must become tools in farmers’ hands, and we will be watching implementation every step of the way.”

Farmers have been fighting against John Deere’s repair practices for more than a decade, but the FTC began its fight in 2021, led by then-chair Lina Khan under the Biden administration. In April, John Deere agreed to pay out $99 million in a separate class action lawsuit brought against the company in 2022. Repair and consumer advocates say this FTC settlement does far more to help farmers than a payout did.

John Deere has maintained that it already has robust repair resources for its customers, including service manuals and diagnostic equipment. In John Deere’s press release, the company says the settlement is in line with what it has been doing all along, saying that “the agreement reinforces Deere’s continued innovation toward more flexible repair options, emphasizing increased access and transparency for customers. It formalizes Deere’s ongoing commitment to expanding access to diagnostic and repair tools.”

The consumer advocacy group US PIRG issued a statement about the settlement, citing the organization’s 2022 official complaint to the FCC about John Deere’s repair policies.

“We should be able to fix our own stuff,” wrote PIRG’s Right to Repair campaign director Nathan Proctor. “This settlement from the FTC gives farmers more and better options to repair their equipment. It is a win for farmers and all of us who want a more fixable world.”

Some material also goes to an unnamed US technology and industrial company, under a deal penned in the first quarter of 2026.

In the same quarter a year ago, the largest portion of MP’s sales by revenue—mined material, not NdPr—went to China’s Shenghe Resources. But MP has stopped selling to Shenghe as part of its deal with the US government.

MP ultimately plans to produce its own magnets at scale, which would require it to consume much of what it produces. Mined rare earths are turned into oxides, which are used to make metals and alloys that go into magnets.

The company has penned agreements with General Motors and Apple to supply them with its magnets. It said in May that it expected to begin shipping finished magnets to GM this year.

Meanwhile, Energy Fuels—which won $725 million in conditional government funding in June—plans to scale its production of rare earths and also has eyes on Asia.

“We will be sending oxides in the near-term to Korea,” said chief executive Ross Bhappu. Last year, a major South Korean manufacturer made a small amount of Energy Fuels’ NdPr into magnets.

Energy Fuels is in the process of acquiring Australian Strategic Materials, which owns a rare earths metal-making plant in South Korea. It also announced a $1.9 billion deal to buy German magnet maker Vacuumschmelze (VAC) in June, which Bhappu said would result in more of Energy Fuels’ products going to VAC’s US operations.

China is the largest global producer of the widely used neodymium iron boron magnets. Outside China, Japan produces 10,000-15,000 tonnes per year, while South Korea produces 2,000-3,000 tonnes annually, and the US produces 1,000 tonnes or less, according to John Ormerod, a rare earths consultant at JOC LLC. There is also some production in Europe.

Phoenix, which secured a conditional $500 million from Washington in June, said government funding would help it scale up metal and oxide production, which would “expand the pie for everyone.”

MP’s recent earnings have been boosted by the money it receives under its US government deal—which guarantees a minimum sale price for some products and tops up any shortfall from the price paid by third parties.

© 2025 The Financial Times Ltd. All rights reserved. Not to be redistributed, copied, or modified in any way.

OpenAI on Wednesday launched GPT-Live, a pair of new voice models that fundamentally redesign how people talk to ChatGPT — replacing the company’s existing Advanced Voice Mode with an architecture that can listen and speak simultaneously, much like an actual human conversation.

The two models, GPT-Live-1 and GPT-Live-1 mini, are rolling out globally starting today across iOS, Android, and ChatGPT.com. GPT-Live-1 becomes the default voice model for paid ChatGPT users on the Go, Plus, and Pro tiers, while GPT-Live-1 mini serves free-tier users. OpenAI also plans to bring the models to the API, and developers can sign up to be notified.

The release marks the third generation of ChatGPT’s voice technology in roughly two years — and OpenAI’s clearest bid yet to turn its chatbot into something that feels less like querying a search engine and more like talking to a colleague.

Why full-duplex voice changes everything about talking to AI

The defining technical advance in GPT-Live is what OpenAI calls a “full-duplex architecture.” In telecommunications, full-duplex means both parties on a phone call can talk and listen at the same time. Applied to AI, it means the model continuously processes your incoming audio even while it generates its own spoken response — no more waiting for a clean silence gap to figure out when you’ve finished a thought.

“Instead of processing a sequence of separate messages, GPT-Live continuously processes input while generating output,” OpenAI wrote in its research blog. “The model can therefore make interaction decisions many times per second: whether to speak, continue listening, pause, interrupt, or invoke a tool.”

In practice, that translates to a voice assistant that can insert conversational acknowledgments — “mhmm,” “yeah,” “got it” — while you’re still talking, pick up on a natural pause without jumping in prematurely, and handle rapid interruptions without derailing the entire exchange. 

OpenAI’s previous Advanced Voice Mode, launched to paid users in September 2024, processed and generated audio within a single model but still operated on rigid turn-by-turn exchanges. As OpenAI acknowledged in the announcement, “because turn detection is based on silence, even a brief pause or background noise could be mistaken for the end of turn — causing the model to interrupt at unnatural times.”

That brittleness created a product that, while impressive in demos, could be deeply frustrating in extended real-world use. Background chatter in a coffee shop could trigger a response. A thinking pause might get swallowed. The experience felt, as one researcher put it on X shortly after the announcement, like “walkie-talkie turn taking.” GPT-Live is designed to end that era.

How OpenAI split voice and intelligence into two separate layers

GPT-Live introduces a second structural change that may prove just as consequential for enterprise adoption: it decouples the voice interaction layer from the reasoning layer.

When a user asks a straightforward question, GPT-Live handles it directly. But when the query demands web search, deeper reasoning, or more complex agentic work, GPT-Live delegates the task to a frontier model running in the background — at launch, GPT-5.5, the large language model OpenAI released in April — and continues talking with the user while the computation happens asynchronously.

“While it works, GPT-Live can keep talking with you and maintain the flow of conversation,” OpenAI explains. “As we release new frontier models, we’ll continuously update the model used by GPT-Live.”

This delegation model is a meaningful architectural bet. Rather than building a single monolithic voice model that tries to be both conversationally fluid and deeply intelligent, OpenAI has split the problem in two: a voice-native model optimized for real-time interaction, and a separate reasoning engine that can be swapped out as the state of the art improves. 

It is, in effect, a modular design — one that allows OpenAI to upgrade the intelligence of its voice assistant without retraining the voice model itself. The implications for enterprise and developer workflows are significant. A voice agent built on this architecture could maintain a natural conversation with a customer while simultaneously querying databases, searching the web, or performing multi-step reasoning — tasks that would have introduced several seconds of dead air under the old pipeline.

The three generations of ChatGPT voice, from clunky pipeline to continuous stream

To understand how far voice AI has come, it helps to trace the three generations that led to GPT-Live.

The original ChatGPT Voice, launched in 2023, used a cascaded pipeline — a speech-to-text model (Whisper) transcribed what you said, a large language model (GPT-4) generated a text response, and a text-to-speech model converted that response back into audio. Each handoff introduced latency and lost information. 

As OpenAI noted, “the complexity came at a cost: information could be lost across models, and responses were slow and stilted.” That cascaded approach was the industry standard, and its limitations were well-documented. As the blog OpenHelm noted in an October 2024 analysis of OpenAI’s Realtime API, the old pipeline stacked up to roughly 1,700 milliseconds of latency — nearly two full seconds of dead air before the first word of a response. Managing the state between the three separate APIs consumed an enormous amount of engineering effort.

OpenAI’s Advanced Voice Mode, which began its limited rollout to paid ChatGPT Plus users in July 2024 before expanding more broadly in September 2024, collapsed that three-model pipeline into a single model that processed audio natively. As TechCrunch reported at the time, the rollout came with five new voices — Arbor, Maple, Sol, Spruce, and Vale — alongside improved accent handling and smoother conversations. 

The feature also launched on the web in November 2024, extending it beyond mobile. But Advanced Voice Mode still operated through discrete, alternating turns — and it launched into the shadow of a PR debacle that OpenAI is still working to leave behind.

The Scarlett Johansson controversy still shadows OpenAI’s voice ambitions

Advanced Voice Mode arrived in the wake of one of OpenAI’s most damaging self-inflicted crises. During the GPT-4o launch in May 2024, the company showcased a voice called “Sky” that many listeners immediately noted sounded strikingly similar to Scarlett Johansson, who famously voiced an AI companion in the 2013 film Her.

Johansson said she had declined OpenAI CEO Sam Altman’s offer to voice the system, then was “shocked, angered and in disbelief” when the product launched with a voice her own friends couldn’t distinguish from hers, as NBC News reported. Altman had tweeted just the word “her” the day the product launched.

OpenAI pulled the voice and apologized, but the incident drew public scrutiny from SAG-AFTRA and members of Congress, and crystallized broader concerns about AI companies moving fast with creative IP.

The Hollywood labor union said the issue underscored “why we’re strongly championing federal legislation that would protect their voices and likenesses … from unauthorized digital replication,” as NBC News reported. Forbes contributor Paul Tassi wrote at the time that Altman, “by holding up Her on a pedestal of something to strive for, has missed the point of that film” — in which the protagonist’s relationship with his AI companion ultimately does him more harm than good.

GPT-Live appears designed, in part, to move past those controversies. OpenAI says it has “remastered the nine distinct voices in ChatGPT for GPT-Live” and notes the system “is designed for conversation, not voice impersonation,” with “safeguards to prevent it from imitating a real person’s voice.”

What 150 million weekly voice users will actually notice today

OpenAI disclosed that more than 150 million people talk to ChatGPT using voice and dictation features each week — a notable slice of the platform’s 900 million total weekly active users. The voice experience has grown into a substantial product in its own right, used for language practice, bedtime stories, commute-time chat, and hands-free everyday help.

The new product features reflect that usage. GPT-Live introduces rich visual cards that surface during voice conversations — weather forecasts, stock data, sports scores, and maps — giving users something to glance at without breaking the flow of speech.

Users can now choose between three reasoning levels for answers: Instant for quick responses, Medium for moderate thinking, and High for more complex work. And if you take a moment to think, “ChatGPT Voice now waits instead of jumping in and interrupting,” OpenAI wrote. “If you ask it to stay quiet and listen, it will. And when there’s background noise, like passing traffic or nearby conversations, ChatGPT is better at focusing on your voice instead of getting distracted.”

Early reactions from users with preview access were cautiously positive. “I had early access to sol. it is a phenomenal model,” wrote one user on X, adding it is “much better at frontend, long context knowledge work, and its vibes are much better.” Another observer cut to the heart of the matter: “The smarts are not new here, GPT-Live hands hard questions to GPT-5.5. What is new is the feel: full-duplex voice that listens while it talks.”

New voice-specific safety tests reveal where the risks still live

The GPT-Live system card, published alongside the announcement, reveals a safety strategy built around the particular risks of real-time voice interaction — a domain where the speed and intimacy of conversation create hazards that text-based chat does not.

OpenAI expanded its safety evaluations to include audio-native tests, using both real user voice samples (from those who opted in) and synthetically generated prompts targeting edge cases across categories like self-harm, sexual content, illicit behavior, emotional reliance, mental health, and hate speech.

On the synthetic evaluations — which OpenAI described as deliberately adversarial — GPT-Live-1 showed substantial improvements over Advanced Voice Mode. In illicit behavior, for instance, the safety score rose from 0.63 to 0.97. On self-harm, it climbed from 0.72 to 0.98. Hate speech achieved a perfect 1.00, up from 0.87.

On the production-prompt evaluations — which used real user audio and reflected more ambiguous, borderline scenarios — the picture was more mixed. GPT-Live-1 matched or improved on Advanced Voice Mode in most categories but showed a slight regression on emotional reliance (from 0.88 to 0.82), though OpenAI noted the change was not statistically significant.

The company built real-time safeguards that can intervene while the model is speaking — steering toward safer responses, surfacing crisis resources, or ending the voice conversation entirely in higher-risk situations. It also designed additional protections for teen users and adapted self-harm support flows for voice, including crisis helpline integration.

Perhaps most notably, OpenAI said it is “rolling out longer-term measurement and post-launch monitoring focused on emotional reliance” — an acknowledgment that the very naturalness GPT-Live strives for creates its own category of risk.

Google, ByteDance, and Nvidia are already in the full-duplex race

While OpenAI was refining its safety guardrails, its rivals were shipping full-duplex systems of their own. Google’s Gemini Live, which supports full-duplex conversation alongside camera and screen sharing — capabilities GPT-Live notably lacks at launch — is already available in the Gemini app. Google released Gemini 3.1 Flash Live in March as its highest-quality real-time audio model, targeting low-latency voice interactions for developers.

ByteDance launched Seeduplex in April, claiming to be the first production-scale full-duplex speech AI deployed at scale, inside its Doubao app. Seeduplex reported roughly a 50 percent reduction in false-response and false-interruption rates compared to ByteDance’s previous half-duplex system. And Nvidia’s PersonaPlex, released in January, brought customizable voice and role control to full-duplex models, breaking what had been a constraint where natural-sounding models were locked into a single fixed voice.

The competitive picture is clear: full-duplex voice interaction is quickly becoming table stakes for consumer AI products, not a differentiator. OpenAI’s advantage lies in the scale of its existing user base, its integration with GPT-5.5’s reasoning capabilities, and the breadth of the ChatGPT ecosystem.

But the window in which any one company has a monopoly on natural-sounding voice AI has already closed. OpenAI also acknowledged several gaps. GPT-Live does not support voice with video or screen sharing at launch. Language support is limited, with the company noting that “for certain languages, the model may have a non-native accent or gaps in fluency.” And API access is not available on day one, meaning enterprise developers cannot yet build on GPT-Live directly — a constraint that will slow the model’s penetration into commercial voice-agent workflows where competitors like Google, ElevenLabs, and Deepgram already have developer-facing products.

The end of the chat box may be closer than anyone expected

GPT-Live is essentially OpenAI’s most significant bet yet on voice as the primary interface for AI — not just a convenience feature bolted onto a text chatbot, but a purpose-built interaction layer that sits between the user and the company’s most powerful models.

“Over time, we believe this research will also unlock the ability to use voice for increasingly complex, longer-running, and more agentic work,” OpenAI wrote. That ambition — using natural voice as the front end for autonomous AI agents that can perform multi-step tasks — is the logical endpoint of the full-duplex plus delegation architecture.

Imagine telling your phone to book a flight, negotiate with your insurance company, or debug a production server, all through a conversation that feels as natural as talking to an assistant who also happens to have the intelligence of a frontier AI model.

Two years ago, talking to ChatGPT meant dictating into a microphone and waiting nearly two seconds for a stilted reply. One year ago, it meant a smoother exchange that still felt like a polite, slightly awkward phone call with someone who insisted on waiting for you to finish every sentence. Today, it means something closer to a real conversation — imperfect, still constrained in some languages and missing video, but unmistakably closer. OpenAI once got into trouble for wanting to recreate the movie Her. With GPT-Live, the company may finally be reckoning with the harder question the film actually posed: not whether AI can sound human enough to talk to, but what happens to us when it does.

SpaceXAI has released its latest model, Grok 4.5 — the first since the company went public several weeks ago.

In a blog post published Wednesday, SpaceXAI characterized its new release as a workhorse that can tackle all of the typical tasks that the AI industry has sought to automate: coding and app-building, office and clerical work, research, writing, and other forms of routine knowledge work.

Grok can supposedly do all this for less spend, too, as SpaceXAI says that its model has “twice greater token efficiency” than other leading models. If it carries through to real-world use cases, that efficiency would be a big advantage for SpaceXAI, since the cost of tokens has been a growing concern for AI consumers.

The company released benchmark metrics Wednesday that appeared to show Grok’s competitiveness with other top models from SpaceXAI competitors, although just short of best-in-class:

In a post on his social media platform X (which is a subsidiary of SpaceXAI), founder Elon Musk compared the model to Opus, Anthropic LLM designed for intensive and complex tasks.

“Based on strong positive feedback from customers in our beta test program, @SpaceXAI will make Grok 4.5 available to the public tomorrow. It is an Opus-class model, but faster, more token-efficient and lower cost,” wrote Musk in a post on X.

Musk later added: “Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster. The combination of capability, faster speed and lower cost is what makes it competitive.”

SpaceXAI says that its new model costs $2 per million input tokens and $6 per million output tokens. That’s quite competitive, if Grok’s capabilities match SpaceXAI’s rhetoric.

Opus 4.7, by comparison, costs $5 per million input tokens, and $25 per million output tokens. OpenAI has tiered costs for different model versions: Sol, its most expensive, costs $5 for input tokens and $30 for output, while its least expensive, Luna, costs $1 for input and $6 for output.

It’s a big week for AI model releases. OpenAI is planning to release GPT 5.6, its latest, most powerful model, on Thursday. The release of that model had previously been limited by the Trump administration, due to concerns about its security implications. OpenAI has called it its “strongest model yet.”

A China-linked threat cluster has been exploiting vulnerable Roundcube servers at U.S. and Canadian universities to steal credentials and deploy backdoor malware.

The campaign has been observed since May and focuses on physics and engineering departments, administrators and professors, as well as organizations involved in astrophysics, particle physics, or national security-related research.

Researchers at cybersecurity company Proofpoint are tracking the activity under the name ‘UNK_MassTraction’ and believe to be associated with a new threat cluster.

The attack begins with a malicious email sent from compromised accounts or spoofed domains, using a generic lure.

Opening the email in a vulnerable Roundcube webmail client triggers exploitation of a cross-site scripting flaw tracked as CVE-2024-42009, which executes JavaScript code inside the victim’s browser, loading a payload called IceCube.

According to the researchers, IceCube “is a fully-featured Roundcube stealer” that can harvest usernames, passwords, cookies, two-factor authentication (2FA) data, and browser information.

Proofpoint says that the malware uses “helpers” to exploit a Roundcube deserialization flaw tracked as CVE-2025-49113 and attempts to install SquareShell, a PHP webshell that includes remote code execution capabilities.

If successful, the attacker gains remote code execution on the mail server; otherwise, the malware downloads a shell script that loads another payload, VShell, directly in memory.

VShell is a commodity Go-based backdoor that supports interactive shell access and port forwarding, which is commonly used by Chinese threat actors.

Based on several observations, Proofpoint assesses that UNK_MassTraction is likely a China-aligned espionage actor.

First, the infrastructure used in the attacks overlaps with a covert VPS network previously associated with multiple China-linked actors. Another clue is the presence of Chinese-language artifacts in earlier phishing emails.

Finally, the tactic of targeting internet-facing mail servers as a foothold for accessing internal networks is a hallmark of Chinese attacks.

Taking everything into account, Proofpoint emphasizes that attribution in this case is just an assessment and definitely not a high-confidence one.

An interesting finding regarding the specific targeting of this campaign is that UNK_MassTraction appears to have selected servers previously deemed vulnerable to CVE-2024-42009 and CVE-2025-49113, so some reconnaissance was performed prior to the attacks.

Administrators of Roundcube systems are advised to apply the latest security updates that address the two flaws and treat mail servers with the same diligence they show for VPNs and other remote access nodes.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Get the whitepaper