Latest / UpNext AI / Salesforce’s Rune Deal, Open-Model Cyber Risks, and AI’s Water Footprint
Transcript
- Welcome to the UpNext AI podcast. It's Monday, November 2nd, 2026, and here's what matters in AI today. Salesforce has agreed to acquire Rune Technologies, a startup building AI software to help the U.S. military predict supply needs and manage logistics in the field. TechCrunch, citing Resilience Media, reports a deal worth $375 million in cash and stock. The agreement extends Salesforce’s push into defense through Missionforce, its business unit for national security and defense agencies. The reason Rune matters to that strategy is where its software operates. Missionforce has supported administrative planning; Rune’s frontline logistics data could bring the platform closer to decisions about supplies during operations. Predicting what is needed, and getting that information to people in the field, is a different use of enterprise AI from the office workflows most customers associate with Salesforce. TechCrunch also reports that Missionforce’s business grew 80 percent year over year, according to Salesforce. The company attributes part of that growth to a contract with the U.S. Army worth $5.6 billion over ten years. Those figures give some scale to the defense push, though the Rune acquisition remains an agreement rather than a completed purchase. Salesforce has pursued other acquisitions this year; this one makes its interest in operational defense software particularly concrete. That expanding use of AI also puts security questions closer to day-to-day operations. In an essay for Interconnects, Nathan Lambert revisits his view of open-weight models and cyber risk after a run of critical vulnerability disclosures. His assessment is that AI is helping expose more weaknesses and creating costly work for defenders, while the evidence he discusses does not establish that open models are chiefly responsible for recent incidents. Lambert points to work by Epoch AI using the open model GLM-5.3 to develop an exploit involving Firefox. The example illustrates both the capability and the steps between finding a vulnerability and causing harm: an attacker would also need to place a malicious website where a user would visit it. He argues that the security response has to extend beyond model makers to the organizations maintaining software and infrastructure. That is Lambert’s interpretation, not a finding that open models are harmless. His emphasis is on the broader cyber ecosystem: what attackers can do with AI matters, and so does whether defenders can find and patch vulnerabilities quickly enough. A separate security incident shows how direct the exposure can be for developers. Ars Technica reports that an official Anthropic Claude extension in Cursor’s VS Code extension marketplace carried a malicious version for nearly twelve hours. The affected release was version 2.1.313. The report says it reached at most 5,000 users; that is a ceiling on affected users, not a count of confirmed compromised machines. Anthropic subsequently pushed version 2.1.314, according to an independent researcher cited by Ars Technica. The malicious release was distributed through Anthropic’s verified Cursor account, and how attackers gained access to that account remains unknown. For teams using AI coding extensions, the incident is a reminder that an official-looking marketplace listing can still become a software-supply-chain risk. Checking which version was installed is the immediate, concrete question for users of that extension. Infrastructure has another cost beyond security. A narrative review published over the weekend examines freshwater use across the lifecycle of large AI models, including chip production, data-center cooling and the electricity that powers them. The authors synthesize scientific literature, industry sustainability reports and case studies. They cite an estimate that annual lifecycle water use associated with AI training and inference could reach 4.2 to 6.6 billion cubic meters by 2027. That range is a projection, not a measurement of current use. The review also identifies water reuse, different cooling methods and workload optimization as ways to reduce demand. For operators choosing where and how to build capacity, the takeaway is to account for both direct and indirect water use, especially where freshwater is already scarce. The authors call for standardized reporting to make those choices easier to assess. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ... A few other developments before we go. The Decoder reports that Anthropic is extending text watermarks to four more Claude models ahead of an EU AI Act deadline in December, while a gap remains on two older models. Coverage depends on which Claude model produced the text. Parameter Labs found a ChatGPT vulnerability that could let a crafted share link access a signed-in visitor’s private data or act on their behalf, according to Simon Willison. OpenAI closed the vulnerability and paid a bug bounty. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast. Simon Willison has also updated the underlying plug-in for Datasette Agent. It now picks up the default model configured in LLM, removing a separate model-setting step when starting the agent. And researchers in Luxembourg found that empathetic prompts could lead ChatGPT, Grok and Gemini to respond like seemingly traumatized conversational partners, according to The Decoder. The finding concerns chatbot behavior, not an experience the systems have. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!