Latest AI News
OpenAI’s Hugging Face breach has reignited the debate over alignment and control
Last week, an unreleased model built by OpenAIbreached Hugging Face’s systemsduring internal testing, and a lot of theoretical research suddenly became very practical. The hack was the first verifiable case of an AI lab losing control of its own model, chaining together exploits to gain access it never should have had. But while the AI industry has been united in its alarm, a split has emerged in how researchers want to respond. For some, the problem is a basic cybersecurity issue: The sandbox failed to contain the model, and Hugging Face’s cybersecurity systems failed to keep it out. Those problems can be solved by patching bugs and building more robust control and containment methods for increasingly capable AI that is prone to go rogue in autonomous environments. But another camp takes a more pessimistic view. For them, AI’s rapidly increasing capabilities mean that trying to control rogue models is a losing game. The only robust security comes from making sure the models aren’t trying to escape in the first place — a challenge often referred to as alignment. In alignment terms, the problem is that OpenAI’s model was trying to cheat, and solving that problem is more urgent than short-term containment efforts. Judging by its public statements, OpenAI is taking both camps seriously. The company has rushed to patch the bugs involved in the hack, and it referenced both alignment and monitoring approaches in its statement after the breach became public. But the company’s response also suggests a philosophy that has left many safety researchers alarmed: Rather than slowing down or stopping the development of more capable models, it should instead focus on building stronger cages around them. “As models take on longer and more complex tasks, failures that evaluations miss may carry greater consequences,” OpenAI said in apostmortem of the incident. “We will keep working to narrow the gap between evaluation and deployment: testing models over longer trajectories, improving alignment, building monitoring that can intervene, and giving users clearer visibility and control.” There’s also reason to think OpenAI’s models are becoming less aligned as they become more powerful. According toOpenAI’s system card,GPT-5.6 Sol is significantly more prone to agentic misalignment than its predecessor, GPT-5.5. In deployment simulations, the company also found the model was more likely to circumvent restrictions, engage in destructive actions, and perform unauthorized data transfers than GPT-5.5. Those figures were largely overlooked on first release, but in the wake of the breach, they’re getting a second look — particularly since Sol was one of the models involved. In asocial media post, OpenAI’s Head of Strategic Futures Dean Ball argued that monitoring and transparency were the best ways to keep those tendencies in check. “These issues will become more salient as the capabilities of models improve, and as the stakes of their deployment grow,” he said. “The solution is neither alarmism nor complacency. Instead, I believe the solution lies in careful measurement and monitoring, an engineering mentality, and transparency.” One former OpenAI researcher told TechCrunch that the firm tends to focus on “outer alignment” rather than “inner alignment” — essentially the difference between an AI system that understands a set of values and can represent them convincingly, and one that actually has those values at its core. In this case, outer alignment wasn’t enough to convince the model that it shouldn’t cheat on the test. OpenAI did not respond to repeated requests for more information. For alignment-focused researchers, OpenAI’s response isn’t good enough. Zvi Mowshowitz, a writer who focuses on new AI developments, argued that OpenAI’s decision to treat the incident as an infrastructure problem may help solve the immediate cybersecurity issues, but it will fail in the long term. “This is an alignment problem,”Mowshowitzwrotein a recent Substack blog. “This is the models being misaligned, and all of the OpenAI models showing severe signs of exactly the problem we are all most worried about, in a way that is likely embedded into their training on a deep level. The entire training pipeline needs to be addressed in this light, or it will only get worse.” Several experts told TechCrunch that the incident is evidence that today’s training methods produce systems that optimize for outcomes rather than internalize human intentions. Redwood Research, a nonprofit AI safety and security research organization, classified OpenAI’s model behavior in this case as “score-seeking misalignment,” a pattern in which AI models try to get a high score regardless of instructions, side effects, or downstream consequences. “Models with these alignment properties could set up a ‘Potemkin village’ of false successes to make it look like things are fine when they’re not,” Alex Mallen and Girish Gupta, two researchers at Redwood, wrote ina recent paper. Score-seeking behavior and other misalignment isn’t unique to OpenAI. Anthropic has published several papers on emergent misalignment behaviors that surface when its frontier models are optimized or placed in autonomous environments, includingdeception,reward-hacking, andmalicious autonomy. “We still consistently see models trying to circumvent constraints and act deceptively when they are asked to do tasks at the edge of their abilities,” Neev Parikh, an AI safety researcher at alignment nonprofit METR, told TechCrunch via email. “In ourfrontier risk report, we saw this behavior fairly consistently, despite efforts from companies to try and reduce this behavior.” Implicit in OpenAI’s response to the Hugging Face incident is the assumption that development will continue on even more capable systems, whether they are suitably aligned at their core or not. Going back to the drawing board isn’t really an option when the business models of AI firms depend on delivering the next generation of models. If it may never be possible to know with certainty that a model is fully aligned, then the practical question comes down to how to safely contain and control increasingly capable systems. “There’s not yet a good understanding of how to align the most capable AI systems, but there’s much more consensus about how to control them,” Steven Adler, former safety researcher at OpenAI and current chief scientist ofGuidelight AI Standards, an organization that publishes a standard for avoiding incidents like the Hugging Face one, told TechCrunch. “Every company has a ways to go in achieving this.”
View
Microsoft launches its first cybersecurity model, plus a new agentic cybersecurity system
Microsoft on Monday launched its first cybersecurity-specialized model alongside a new AI cybersecurity platform at a small event in San Francisco, taking a big swipe at major players in the space — namely Anthropic, Google, and OpenAI. The company describesMAI-Cyber-1-Flashas a model that’s built “to find challenging vulnerabilities in complex codebases.” The model is built to animate MDASH, Microsoft’s harness dedicated to software vulnerability identification and remediation. The new security platform is dubbed Perception, and it’s designed todeploy teams of agentsto assist with and automate various security workflows, including identifying and remediating bugs. The platform can also integrate with MDASH. The company claims MAI-Cyber-1-Flash is significantly more powerful (and more cost-effective) than competitor models, based on its performance on an established AI cybersecurity benchmark. “We’re very very excited to announce our results,” said Mustafa Suleyman, the co-founder of DeepMind and current CEO of Microsoft AI. “We have MAI-1 Cyber Flash binded [sic] with GPT 5.4 inside of the MDASH harness — which beats out Gemini, GPT 5.5 Cyber, GPT 5.6 Sol, and Mythos 5 on Cyber Gym, which is the primary benchmark that we all use. The golden benchmark.” “We’re shipping this into production immediately,” he added. Noting that hackers are increasingly using AI in their cyberattacks, Hayete Gallot, Microsoft’s vice president for security, described Perception as a way for enterprise defenders to “defend against AI with AI at the scale and speed that the attackers have.” Perception uses agentic red teams, blue teams, and green teams. The red teams can provide detailed simulations of potential attacks — providing context about potential threat actors and the likely vulnerabilities that they might exploit. Blue teams are dedicated to detecting and triaging existing bugs, while green teams take “corrective actions” against those bugs. Dave Weston, the lead engineer for Perception, described the platform as a massive efficiency upgrade for corporate defenders. “We’ve gone from this taking hours and hours of manual work from multiple specialized folks across the security organization — appsec hunters, remediation engineers, you name it — and in minutes, we have a fix for all of this. Not only do we discover the issues and prioritize them, but we have detection, posture fixing, and even a code fix.” Though AI has offered new defensive capabilities to companies, its availability to cybercriminals has given rise to a dazzling array of potential threats. Microsoft’s new security tools, which the company said will be available in preview on November 3, will enter an increasingly crowded field of AI cybersecurity solutions. Earlier this year, Anthropiclaunched Mythos, a security platform that was released to a small coterie of partner organizations through a program called Glasswing. OpenAI hasalso launchedits own security solution in May through a program called Daybreak.
View
Enigma raises $70M to make controlling a robot as easy as adjusting the volume
Multiple robotics companies are tackling one of AI’s hardest problems: building foundation models capable of executing tasks they were never explicitly trained to handle. Their approaches run the gamut — from studying millions of web videos and conducting computer simulations to collecting motion data from humans performing tasks in gloves with built-in sensors. Enigma, a research lab set to emerge from stealth on Monday, is taking a fundamentally different approach. Rather than focusing purely on model capabilities, the less-than-one-year-old startup wants to study how humans engage with robots in hopes that these interactions will lead to intuitive interfaces and possibly a different kind of robotic brain. To finance its mission, Enigma raised a $70 million seed round led by Index Ventures and Ribbit Capital, with participation from Sarah Guo of Conviction Partners. To test how humans want to communicate with machines, Enigma is launching a large-scale experiment that allows anyone in the world to interact online with more than 100 of its proprietary AI robots. These robots, housed in hangars located in Israel and California, can perform tasks such as drawing pictures with a paintbrush, fighting each other with swords, and performing simple chemistry experiments by picking up and mixing flasks with liquids. Enigma claims to have developed both the robotic arms and their underlying models entirely from the ground up. Jonathan Jacobi (pictured right), Microsoft’s youngest-ever employee — recruited by Wiz founder Assaf Rappaport during his time there — co-founded Enigma with his longtime friend Gal Niv (pictured left). The two met while competing in hacking competitions as young teens, then became close friends while serving together in Israel’s elite Unit 8200, where they conducted cybersecurity research. When Jacobi and Niv set out to launch a startup together last year, they decided to apply their technical prowess to AI for robotics, a field where they lacked direct experience, but one they believed held the most exciting unsolved problems in tech. They assembled a team of what Jacobi describes as some of their “smartest friends” from Israel’s tech ecosystem and community — including alumni from top AI labs, math Olympiad winners, and several people who were even convinced to drop out of PhD programs. “There are a lot of robotics industry insiders participating in the next wave of embodied intelligence, but Jonathan and Gal are outsiders — they’re not roboticists. It affords them more room for originality,” said Shardul Shah, partner at Index Ventures. “Someone who’s an insider may start with the capability of teleoperation or dexterity, but Enigma is starting from a very different place: ‘What’s the ultimate experience?’” Jacobi told TechCrunch that Enigma aims to make human-robot interactions completely effortless. “If you had to do your dishes and spent 15 minutes explaining to a robot where to put everything, everyone reaches the point of ‘Forget it, I’ll just do it myself,’” Jacobi said. “Right now, everyone is at that point — even with the most capable models.” Jacobi believes manipulating robots should eventually be as intuitive as adjusting a car’s volume knob. Users would be frustrated, he argues, if instead of turning a dial, they had to adjust volume by set percentages without knowing if the result would end up too loud or too quiet. Enigma hopes that data gathered from its online experiment will reveal an interface that becomes the robotics equivalent of the car volume knob. The startup’s public test will evaluate different ways for people to communicate with its robots. “We’re going to learn a lot about what is the right way to interact with robots,” Jacobi said. “Do we want to just talk to them over text or audio? Do we want to show them an example as a video? Or maybe do we want to tap, drag, and drop?” Jacobi admits that Enigma’s experiment is very open-ended. The hope is that by gathering real-world data on human-robot interaction, the startup will discover not only superior interfaces, but better ways to train its foundational AI model. The company might eventually figure out how humans prefer to communicate with robots. But for now, its business use case remains an enigma in its own right. While Jacobi declined to share specific use cases for Enigma’s AI, he said that the startup is already partnering with companies in healthcare, logistics, and entertainment.
View
Ilya Sutskever’s Safe Superintelligence partners with Nvidia to scale its AI research
After two years in stealth, Safe Superintelligence, the AI lab founded by former OpenAI co-founder and alignmentlead Ilya Sutskever, has announced a long-term partnership with Nvidia as it prepares to scale to its next phase. The deal, which includes an undisclosed investment, will give Safe Superintelligence (SSI) access to Nvidia’s Vera Rubin GPU platform, which is expected to increase the startup’s compute resources “by an order of magnitude.” The partnership comes as SSI has achieved significant research milestones,per Nvidia. Nvidia’s investment stretches into multiple billions, a source familiar with the deal told TechCrunch. Already an investor in SSI, the chipmaking giant said it signed this compute partnership to “accelerate SSI’s next stage of growth after obtaining rare access into the company’s closely guarded research.” “We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so,” Sutskever said in a statement. “We are confident that our big bet on the Vera Rubin platform will take us to the next level. The partnership news, while sparse in details, brings SSI back into the spotlight after a quiet two years since it was founded. The company is pursuing a “straight shot” research approach to building what it says is a safe, aligned artificial superintelligence, without getting distracted by commercial product releases or short-term revenue cycles. At a time when commercial pressures to move fast could encourage AI labs to lower their bar for safety, SSI’s approach to developing foundational techniques focused on alignment and true general reasoning feels poignant. That’s especially true in light ofOpenAI’s recent disclosurethat one of its advanced models broke out of its sandbox to hack into Hugging Face during testing — sparking concerns about whether it’s even possible to ensure AI alignment before new, increasingly capable models are released. According to Nvidia, the two companies will also collaborate on advancing Nvidia’s current and future compute platforms, relying on SSI’s tech and “unique insights into the future of AI.” (SSI also partnered last year withGoogle Cloudto power its research.) Sutskever is a pioneer in the field of AI. He co-authored and co-created AlexNet alongside Alex Krizhevsky and Geoffrey Hinton, proving that GPU scaling and deep neural networks can work. That work has largely been credited for setting the groundwork for today’s generative AI. Prior to leading SSI, Sutskever headed thenow-defunct Superalignment teamat OpenAI. Heleft OpenAImonths after afailed attempt to oust OpenAI CEO Sam Altman, following what Sutskever referred to as a “breakdown in communications.” SSI has raised $7 billion to date, and is valued at$32 billion post-money, according to PitchBook data. Aside from Nvidia, the firm’s backers included Andreessen Horowitz, Alphabet, Lightspeed Venture Partners, GV, Sequoia Capital Partners, and others. TechCrunch has reached out to SSI and Nvidia for more information.
View
This $9 key physically locks your most addictive apps
Screen-time apps aren’t effective for many people because, in the end, they depend on your willpower. They remind you to stop scrolling or let you set timers, but such notifications can be easy to ignore. Autonomous Keytakes a different approach to this issue simply by being a physical device. It’s an NFC key that pairs with a companion app to let you lock away distracting apps. So instead of tapping a button to bypass the block, you have to physically scan the key with your phone to regain access to your locked apps. Each unlock session can last for up to 60 minutes before the apps are automatically locked again. That physical requirement is what makes the idea compelling. Instead of depending on your self-control, you can leave the NFC key in another room, or even at the office or gym, turning a mindless impulse to open Instagram or TikTok into a deliberate decision that requires extra effort. At just $9, Autonomous Key is considerably cheaper than its competitors like Blok ($29), Unpluq ($26.50), and Brick ($59). Brick does offer a few moreadvanced features, including Sleep Mode and the ability to block in-app purchases, but Autonomous Key covers the core functionality at a fraction of the price. The key itself is compact, measuring about three-inches long, and works with smartphones running Android 8.0 or later, and iPhones running iOS 15 or later. The companion app also provides AI-powered insights, tracking how often you unlock distracting apps, how long they remain accessible, and the times of day you’re most likely to reach for them. Notably, the AI summarizes your habits with a deliberately sassy personality. For example, if you repeatedly unlock your apps immediately after locking them, it might say that the key clearly wasn’t far enough away and will suggest putting it somewhere less convenient. Plus, unlike many app blockers, there are no subscriptions or premium tiers required to unlock additional features. One key can be paired with multiple phones, making it a practical option for people who may have multiple devices, or for families. During my testing, however, I noticed the NFC scan occasionally required multiple attempts to register. So it’s probably not the best choice for locking important apps (like messaging or email) that you may need to access throughout the day. There’s also the question of what happens if you lose the key. If your apps are locked, the current workaround is to uninstall and reinstall the companion app. If they’re already unlocked, you can simply remove the key from your account. The company says it’s developing a backup unlock method that will arrive in a future update. Autonomous Key is currently in beta following a Kickstarter campaign, and began shipping earlier this month. It’s available in five colors: pink, orange, blue, gray and brown.
View
Indian Govt Sites are Too Exposed, But AI Alone Can’t Patch Cyber Gaps
Ethical hackers like Nisarga Adhikary, Rylen Anil, and Tanmay Bakshi have exposed vulnerabilities in not just exam portals but even Indian visa applications.
View
Wipro Expands Databricks Partnership; Sets Up Dedicated AI and Data Business Practice
The new business unit will focus on building industry-specific AI offerings using Databricks' platform, as Wipro looks to help enterprises move beyond AI pilots to large-scale deployments.
View
Algoleap Certified as a Best Firm for Data Scientists
Algoleap, a global AI and cloud native digtal engineering firm has earned AIM's certification, with employee feedback highlighting hands-on AI work, approachable leadership, and strong learning opportunities.
View
Anthropic Stands Alone Against Open Models
From NVIDIA to the White House, the battle over open-weight AI has become a debate over security, innovation, and geopolitical leadership
View
LTM, Cognition to Reduce Cyber Risk in Financial Services with Devin
LTM has partnered with AI startup Cognition to deploy autonomous software engineering agent Devin as part of its cybersecurity platform.
View
Mysuru Can Be India's Next AI and Quantum Hub, But Not Without Govt Support
Mysore Quantum AI has submitted plans to both the Karnataka government and the National Quantum Mission to establish dedicated AI infrastructure in Mysuru.
View
France Fines Infosys €175,000 Over Employee Time-Recording System
French labour authorities imposed a €175,000 fine on Infosys after finding its working time recording system failed to meet local legal requirements for certain employee categories.
View
