arXiv:2604.19844v1 Announce Type: new Abstract: Recent advances in embodied Vision-Language Agentic Systems (VLAS), powered by large vision-language models (LVLMs), enable AI systems to perceive and reason over real-world scenes. Within this context, environmental signals such as traffic lights are essential in-band signals that can and should influence agent behavior. However, similar signals could also be crafted to operate as misleading visual injections, overriding user intent and posing security risks. This duality creates a fundamental challenge: agents must respond to legitimate environmental cues while remaining robust to misleading ones. We refer to this tension as trust boundary confusion. To study this behavior, we design a dual-intent dataset and evaluation framework, through which we show that current LVLM-based agents fail to reliably balance this trade-off, either ignoring useful signals or following harmful ones. We systematically evaluate 7 LVLM agents across multiple embodied settings under both structure-based and noise-based visual injections. To address these vulnerabilities, we propose a multi-agent defense framework that separates perception from decision-making to dynamically assess the reliability of visual inputs. Our approach significantly reduces misleading behaviors while preserving correct responses and provides robustness guarantees under adversarial perturbations. The code of the evaluation framework and artifacts are made available at https://anonymous.4open.science/r/Visual-Prompt-Inject.

Tesla expects to start production of Optimus, its humanoid robot, later this year. Costfoto/NurPhoto via Getty Images Tesla hopes to unveil the latest version of its Optimus robot around late summer of this year. Elon Musk said he wants to push the date closer to the start of production because of competitors. Tesla said it's making two factory lines in California and Texas for Optimus production. Elon Musk wants to show off the goods — but not to nosy rivals. During Tesla's Q1 earnings call on Wednesday, investors asked when they could expect to see Optimus V3, the company's humanoid robot, which Musk said will be "Tesla's biggest product ever." The CEO said retooling Tesla's Model S and Model X factory line in Fremont to start production for Optimus is a massive undertaking, and said he also needs to be wary of watchful competitors. "We're also a little hesitant to show V3 off because we find our competitors do a frame-by-frame analysis whenever we release something and copy everything they possibly can," Musk said. "So I think there's some value to not showing new technology until it's close to production." Musk said Tesla expects to start Optimus production in Fremont around late July or August. The CEO didn't name the copycat competitors. Tesla said in a shareholder deck that progress for Optimus production is shaping up as it prepares to build two factory lines — one in Fremont, California, and another at Tesla's Gigafactory in Austin. The Fremont line is expected to produce one million robots a year, while the Texas line will have a long-term annual production output of 10 million robots, according to the shareholder deck. Musk said the Texas factory will likely begin producing Optimus around the summer of 2027. Earlier in January, Tesla announced it would discontinue the Model S and Model X to make room for Optimus production at the Fremont factory. Musk said testing is ramping up and that Optimus could complete tasks outside Tesla's factories "sometime next year." However, the CEO — known for his reach-for-the-sky timelines and projects — wanted to "inject some reality" into some of the questions raised on Wednesday about Optimus. He said dismantling the production lines for Model S and Model X in Fremont will take "at least a few months," and installing a new line will take an additional "several months." "Frankly, if we're able to go from stopping production on one line, dismantling that entire line, reinstalling a whole new line, and turning that on in a matter of four months, that is an insanely fast speed," Musk said, adding that he doesn't know what the production rate of Optimus will be for 2026. Musk also provided a little more color around Optimus' current capabilities. The CEO said Optimus will start with simple factory tasks and that he believes Tesla will be able to pack "a lot of intelligence" into the robot locally — meaning Optimus should be able to perform some tasks without an internet connection. Optimus will still need someone to tell it what to do, Musk said, using the analogy of a manager. He said the robot will need an "orchestration AI" that will guide it, which is where Grok, the AI chatbot made by xAI, will come in. The amount of interaction Grok will have with Optimus could be minimal, Musk said. "Grok would probably have about as much interaction with Optimus as a manager would have with the people on their team," he said. "Meaning Optimus could probably work for several hours without any management oversight." Read the original article on Business Insider

The Tesla Roadster has faced years of delays. Tesla Elon Musk has delayed the Tesla Roadster release once again. Musk teased a potential unveiling of the Roadster in "a month or so." Tesla's Roadster has faced multiple delays after reservations first opened in 2017. Elon Musk has pushed back the release date for the long-anticipated Tesla Roadster — and offered another vague timeline for its release. On Tesla's first-quarter earnings call Wednesday, the CEO said the company might unveil its next-generation Roadster "in a month or so." "It requires a lot of testing and validation before we can actually have a demo and not have something go wrong with the demo," Musk said. "But I think it will be one of the most exciting product unveils ever." Musk said eventually the Roadster would be the only human-driven car in Tesla's lineup, adding that while he doesn't expect it to have a huge impact on the company's revenue, "it is very cool." The Tesla Roadster, an electric sports car, has faced years of delays since it was initially made available for reservations in 2017. Musk said in October the car could be unveiled before the end of the year, only to say a month later that the demo day was tentatively planned for April 1, 2026. In mid-March, Musk said it would "probably" come in late April, only to delay it once again. Tesla beat earnings expectations on Wednesday, signaling a rebound in demand in North America. The company reported a 16% year-over-year revenue increase, beating analysts' expectations. Tesla also said it was planning a large increase in capital expenditures. The stock briefly surged 4% in after-hours trading following the earnings report, but later dropped back down to about what it was at market close. Read the original article on Business Insider

Dario Amodei is the co-founder and CEO of Anthropic, Bloomberg/Getty Images Anthropic's valuation now hovers at $1 trillion on Forge Global, a leading private marketplace exchange. That is a higher price than OpenAI, which was valued at $852 billion earlier this year. Those lucky enough to own Anthropic shares are getting hounded with multiple offers a day to sell. Desperate buyers are in a race to secure a dwindling supply of secondary shares in Anthropic, driving the AI company's valuation on some sites to $1 trillion, a price that would have seemed unthinkable even a few weeks ago. Meanwhile, traders Business Insider spoke with are seeing slumping demand for OpenAI, which is now trading at a discount to Anthropic, despite OpenAI being valued at $852 billion, more than twice Anthropic's valuation in their most recent funding rounds. Anthropic's valuation now hovers at around $1 trillion on Forge Global, a leading private marketplace exchange, its CEO Kelly Rodriques told Business Insider. OpenAI's valuation on the platform is $880 billion, a slight uptick from its March funding round. Since Anthropic and OpenAI are not yet public companies, the vast majority of investors are forced to buy via secondary markets, with existing stock in the companies sold by current or former employees or early investors. Neither company responded to a request for comment. One Anthropic shareholder recently offered to unload shares at a $1.15 trillion valuation, according to Ken Sawyer, cofounder and managing partner at Saints Capital, a venture secondary firm. A "very well known growth fund" offered to buy Anthropic shares at a $1.05 trillion valuation, Jesse Leimgruber, founder of OpenHome, posted on X this week. "Absolutely wild," he said. Some interested buyers have gotten more creative, offering to sell their home in exchange for Anthropic shares at a valuation above $800 billion. A feverish demand for shares It was just three months ago when Anthropic closed a funding round led by GIC and Coatue, valuing the company at $380 billion. Since then, a feverish demand has overtaken Silicon Valley for shares in Anthropic, as investors have been wowed by its torrid revenue growth and momentum around its AI-powered coding assistant, Claude code. "It's been an epic run for Anthropic," said Glen Anderson, CEO of Rainmaker Securities, a merchant bank focused on private securities transactions. "Everybody wants to be part of a generational opportunity in AI, and right now, Anthropic is in the pole position." The company has fielded multiple offers from VCs valuing it at as much as $800 billion in recent weeks, Business Insider reported last week. Anderson just received an offer to buy shares in Anthropic at a $960 billion valuation, a price he says would have been unthinkable even a few weeks ago. But before he can even evaluate the deal, he expects it to be snapped up by someone else. "We get an offer, and then within a day someone else has already bought it," he said. "There are almost no sellers." Those fortunate enough to own Anthropic shares say they are getting hounded with multiple offers a day to sell. "We receive daily offers from the ridiculous to the sublime," said Bradley Horowitz, a general partner at Wisdom Ventures, which was an early investor in both Anthropic and OpenAI. "I barely open those emails because we're not interested. We are playing a long game." Much of the demand is driven by FOMO more than market fundamentals, with investors at venture firms and family offices feeling like they need to own Anthropic shares no matter the price, according to Anderson. "It's almost less about the return than being about to say they're an Anthropic investor," he said. "That drives up the price." Meanwhile, Anderson has seen little demand for OpenAI shares this year, with bids lower than its last round of $852 billion. "OpenAI has been a very tepid market," he said. "The sentiment has certainly shifted to Anthropic. Read the original article on Business Insider

Tesla's planned capex for 2026 is three times higher than what the company has historically spent. Its CFO said, as a result, Tesla will have a negative free cash flow the rest of the year.

In a previous post, we introduced the Universal Sparse Tensor (UST), enabling developers to decouple a tensor’s sparsity from its memory layout for greater... In a previous post, we introduced the Universal Sparse Tensor (UST), enabling developers to decouple a tensor’s sparsity from its memory layout for greater flexibility and performance. We’re excited to announce the integration of the UST into nvmath-python v0.9.0 to accelerate sparse scientific and deep learning applications. This post provides a walkthrough of key UST features… Source

Google has introduced a host of new automated functions into Workspace, all of which are driven by Workspace Intelligence, its new AI system.

Michael Truell's startup, Cursor, is partnering with Elon Musk's SpaceX. Isaac Wasserman/Big Event Media/Getty Images SpaceX said on Tuesday that it's partnering with Cursor, an AI coding startup. The deal gives Cursor access to SpaceX's computing resources. Here's what smart people in business and tech are saying about the multibillion-dollar deal. SpaceX's multibillion-dollar deal with AI coding startup Cursor marks a colossal step for Elon Musk and his ambitions to win the AI race. The space company, which owns AI startup xAi, said in an X post on Tuesday that it's working with Cursor to "create the world's best coding and knowledge work AI." It's expected to combine Cursor's AI-powered coding model with SpaceX's Colossus training supercomputer, the companies said. As part of the deal, SpaceX gains the right to acquire Cursor later this year for $60 billion, or pay Cursor $10 million for the work they produce together. The companies have also held discussions in recent weeks with Mistral about a potential three-way partnership, Business Insider reported. The partnerships could be a crucial step for Musk to get ahead of rivals like OpenAI and Anthropic. Musk, who founded SpaceX in 2002, has made serious strides in AI this year. Chief among them is SpaceX's February acquisition of Musk's xAI, which helped SpaceX expand into AI infrastructure and software. In April, SpaceX confidentially filed for an IPO, boosting gains in other space stocks and teasing a public debut this year. News of the Cursor deal electrified chatter among business and tech industry professionals on social media, many of whom viewed it as a symbiotic match. Here's what people in tech are saying about the deal. Alex Finn, founder of Creator Buddy and Henry Intelligent Machines Alex Finn, founder of Creator Buddy and AI agent startup Henry Intelligent Machines, said the deal made "so much sense" in an X post. "xAI has been behind on coding products for years now. Cursor has a great coding product, but will fail unless they build their own model," he said on Tuesday. The deal could allow the two companies to address those issues, Finn said. While SpaceX gains Cursor's coding capabilities, Cursor gets access to SpaceX's compute infrastructure to build its own model rather than relying on OpenAI's ChatGPT or Anthropic's Claude. Finn said the hurdle Cursor faced is likely happening to many vibe-coding tools that rely on OpenAI and Anthropic, which are building competing offerings. "Wins for both sides," Finn said. Hadley Harris, cofounder of Eniac Ventures Hadley Harris, cofounder of seed-stage VC firm Eniac Ventures, said on X that he didn't "get" the deal. "Every frontier dev I know has moved off Cursor and off IDEs entirely," Harris said on Wednesday, referring to "integrated development environments," or applications that combine building, editing, testing, and other coding capabilities. "Only laggards are still on it. And dev tools always move from thought leaders to laggards, never the reverse." Mario Nawfal, founder of the IBC Group Mario Nawfal, founder of the startup incubator and accelerator IBC Group, said Cursor's users are largely "elite software engineers," who are an important group for SpaceX to cultivate ahead of an IPO. Bringing in more software engineers would help SpaceX, and by extension, xAI, delve further into AI infrastructure and development. "@elonmusk now has space, satellites, AI, social media, and the world's most popular coding tool under one roof," Nawful said. "What he's cooking up will be wild." Tomasz Tunguz, founder and general partner at Theory Ventures Tomasz Tunguz, general partner at early-stage VC firm Theory Ventures, said the partnership allows SpaceX and Cursor to fill their individual infrastructure gaps. "Winning in agentic coding requires three layers: compute, models, & distribution," Tunguz said in an X post on Wednesday. "Anthropic, OpenAI, & Google own the full stack. xAI & Cursor each have gaps," He said xAI has massive compute power, referencing Musk's Colossus data center in Memphis, but the company is losing popularity. Cursor, he said, has the opposite problem. "Millions of developers vibe coding, but its model layer depends on OpenAI, Google, & Anthropic — all competitors. This relationship also pressures margins," Tunguz said. "For $10 billion, SpaceX buys a call option on the distribution it couldn't retain, & Cursor wins the independence it hasn't yet secured." Sarah Catanzaro, general partner at Amplify Amplify general partner Sarah Catanzaro reacted to the deal by referencing Elon Musk's ambitious plan to put data centers in space. Amplify is a VC focused on early-stage tech startups. "I guess Elon realized to get data centers in space, you first need a really good coding agent …" Cantanzaro said in an X post on Tuesday. Anand Kannappan, a former data scientist at Meta and cofounder of PatronusAI Anand Kannappan, cofounder of Patronus AI startup, said on Tuesday that the deal wasn't so much a merger-and-acquisition as it is a "a bet on what the real bottleneck in frontier coding models is." "The deal lets Cursor train Composer on Colossus while xAI runs the same recipe on Grok," its AI assistant, Kannappan, a former data scientist at Meta, said on X. "Both sides find out, at the same time, whether Cursor's data is actually the difference." He added: "The option structure reflects that uncertainty. If the training work ports over, SpaceX buys Cursor and owns the pipeline. If it doesn't, they pay $10B for the experiment and walk." Regardless of the outcome, Kannappan said Musk's companies benefit. "Either outcome, Grok ends up stronger than it would have been, and xAI gets an answer to a question it couldn't answer internally," he said. Aadit Sheth, cofounder of The Narrative Company On Wednesday, a cofounder of The Narrative Company — a communications firm aimed at company executives — said SpaceX's deal with Cursor puts the companies in direct competition with industry leaders like OpenAI and Anthropic. Aadit Sheth said the companies are betting that Musk's AI supercomputer can train a Cursor model that could replace Claude and GPT. Both Anthropic and OpenAI are working to build their own integrated development environments, which help streamline software production. "Cursor has the user. It doesn't have the model. Distribution without a defensible model underneath is a rental," Sheth said in an X post on Wednesday. "We'll know in 6-12 months whether that $60B bought a moat or a rental." Art Levy, chief business officer at Brex Art Levy, chief business officer at fintech company Brex, said he liked the deal in a X post on Tuesday. "This is a 'try before you buy' for Elon, with a massive 'break up fee' for Cursor if it doesn't work out," Levy said. He pointed out that the deal's structure gives SpaceX a call option and prevents "startup destruction" if the deal falls through. "I like it," he said. Max Kolysh, cofounder of Dover Max Kolysh, cofounder of recruiting startup Dover, said Cursor's decision to partner with SpaceX is likely a survival move. He said on X that Cursor's long-term viability had been contingent on its access to Anthropic's and OpenAI's models "Both are actively building Cursor competitors," Kolysh said. "That's an existential platform risk to survive." Kolysh also said Cursor needs its "own foundation models" — like Anthropic's Claude or Google's Gemini. Training those requires deep pockets. "They found the guy with the deepest pockets in the world," he said. Rohit Mittal, cofounder and CEO of Helium Ventures On Tuesday, Helium Ventures CEO and cofounder Rohit Mittal said the deal could stir up the AI startup scene. Helium Ventures acquires and guides software businesses. "It will be very interesting if Claude's token consumption from Cursor moves to xAI," he said on X. Cursor currently relies on Claude, using its tokens — units of data processed by AI models — in its offerings. A new partnership could instead bring xAI into focus, Mittal said. "I can imagine that impacts the growth rate of Claude (not saying it'll slow down), but it'll pull xAI ahead much faster." "The Hunger Games have just begun," he said. Read the original article on Business Insider

Qwen: Alibaba launches Qwen3.6-27B, an open-weight dense model with 27B parameters, saying it surpasses Qwen3.5-397B-A17B on major coding benchmarks  —  · 4226 words  · QwenTeam丨Translations:.体中文  —  HUGGING FACE  —  MODELSCOPE  —  DISCORD

The admission could open Tesla to legal challenges after it spent years promising customers they were just one software update away from owning fully autonomous cars.

X's AI-powered custom timelines are replacing Communities, with Grok-curated feeds...and new ad slots.

“To be clear, we were never approached … nor were we in any discussions regarding a partnership or tour access,” a spokesperson for the artist tells WIRED.

Bloomberg: Dune Analytics: Polymarket's global trading volumes have fallen behind Kalshi in recent months; sources: product delays are contributing to Polymarket's decline  —  Polymarket, the long-time leader in prediction market trading volume, has fallen behind its chief rival as it faces a growing list …

Ship attacked by Iran after possibly falling for safe passage crypto scam.

Disney employees aren't shying away from AI under CEO Josh D'Amaro. Thibault Camus / POOL / AFP via Getty Images; Samuel Boivin/NurPhoto via Getty Images Disney is letting some employees see who the AI power-users are. The company has an "AI Adoption Dashboard" showing who is using the most tokens. One Claude super-fan invoked the chatbot about 460,000 times in a nine-day span. Disney is pulling back the curtain on internal AI usage and letting its employees see who's tokenmaxxing. The Mouse House has given some tech staffers access to an "AI Adoption Dashboard" that tracks token usage across AI coding tools Cursor and Claude, Business Insider has learned. The dashboard shows staffers a wealth of data — including the number of employees actively using AI, the number of requests made, and the number of tokens used over a given period. The dashboard also displays the most active AI users by requests made and tokens used, according to a screenshot of the dashboard viewed by Business Insider. One streaming tech staffer referred to it as a "leaderboard." The biggest AI power users make hundreds of requests per day on average and use tens of millions of tokens, the dashboard shows. Tracking AI usage has led to a so-called "tokenmaxxing" trend, where software engineers compete with their peers to maximize their AI token usage when using chatbots. However, there's debate in the tech world about whether to incentivize high token usage given the associated costs. A person familiar with Disney's approach to the dashboard said it wasn't designed to incentivize high usage but rather to help employees use resources efficiently and effectively. Disney employees are being emboldened by managers to use AI tools more often, the streaming tech staffer said. "They're celebrating it now, but we'll see how long that lasts," this person said of AI usage at Disney. One Disney employee invoked Claude about 460,000 times over nine work days in mid-April, according to the dashboard. That works out to around 51,000 times per day. That high usage likely comes from autonomous agents, a second streaming tech staffer familiar with the AI adoption dashboard said. While this employee said they monitor their AI usage and try to "minimize wasteful token use," they're not worried about it, saying that Disney will "just give me more quota if I get capped." Disney's evolving AI strategy Disney employees aren't the first to get access to an AI usage dashboard. Meta employees created a similar internal dashboard to track Claude token usage, nicknamed "Claudeonomics," before the company shut it down in early April, The Information reported, adding that Meta had used 60 trillion tokens in 30 days. Meanwhile, Visa is encouraging employees to use AI and giving prizes to power users. The payments giant was using 1.9 trillion tokens a month, as of March, the company told Business Insider. Disney's AI adoption dashboard has been around for months, the two streaming tech employees said, so it predates Josh D'Amaro's appointment as CEO. D'Amaro took the top job in mid-March after running Disney's parks and cruises business. D'Amaro must decide what Disney's AI strategy will look like now that the company's blockbuster partnership with OpenAI fell apart. The Disney-OpenAI deal would have given employees access to ChatGPT while putting AI-generated short-form video on Disney+. The Mouse House has warmed up to AI in the last year, several staffers previously told Business Insider, by providing employees with tools like Claude, Cursor, and a "DisneyGPT" chatbot for internal requests. AI is a "top priority" for Disney, a high-level employee with direct knowledge of the company's AI strategy previously told Business Insider. Read the original article on Business Insider

Mark Moran, an underdog Senate candidate from Virginia, claims he wanted to get caught violating the prediction market platform’s rules.

Ryan Roslansky has stepped down as LinkedIn's CEO after six years running the world's largest professional network. Dan Shapero, the company's COO, takes over immediately.

Dan Mangan / CNBC: Kalshi suspends and fines congressional candidates Mark Moran of Virginia, Matt Klein of Minnesota, and Ezekiel Enriquez of Texas for political insider trading  —  Prediction market platform Kalshi said Wednesday it had suspended and fined three candidates for Congress from Minnesota …

IntroductionIn the Databricks intelligence platform, we regularly explore and use...