Latest / UpNext AI / Sophos and OpenAI Daybreak, Pollo Agent, and Anthropic’s OSS Scanner
Transcript
- Welcome to the UpNext AI podcast. It's Friday, October 9th, 2026, and here's what matters in AI today. Security teams need to investigate alerts quickly without handing consequential decisions to an unchecked agent. In a new case study, OpenAI describes how Sophos is using its Daybreak program to bring AI agents into managed detection and response. The agents assemble customer context, detections and threat intelligence, then plan an investigation and draft recommended actions for analysts to review. OpenAI says the average response time for cases using those agents fell from about 38 minutes per case to about 89 seconds per case—a reduction of roughly 96 percent. It also says Sophos now resolves 52 percent of its managed detection and response cases end-to-end with AI, within boundaries set by its analysts. Those are outcomes from the companies’ deployment, not a promise that every security team will see the same result. The operating boundaries matter as much as the speed. Sophos customers can choose whether they act on a recommendation, collaborate on a response, or authorize Sophos to act for them. The company says potentially destructive actions still receive the appropriate human oversight, and cases an agent should not handle go to a person. For security leaders, the useful model here is not simply a faster chatbot. It is an investigation workflow that connects automation to customer permissions and a route for escalation. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast. That same emphasis on workflow shows up in a very different setting: making visual campaigns. OpenAI’s latest Pollo AI case study describes Pollo Agent, a guided video-creation tool that takes a rough idea or visual reference and works it into a storyline, scenes and a script. The platform also helps creators produce images and short video ads from prompts or product photos. Pollo uses different OpenAI models for different jobs. GPT-5.6 routes tasks, GPT-6 Astra handles more demanding narrative and scene planning, and GPT-Image-2.5 generates images. Pollo reports that users spend more than 50 percent less time choosing or switching models. That is a company-reported workflow measure, not a measure of ad performance. The distinction matters for marketers deciding where such a tool fits. Pollo’s pitch is to reduce the work between an idea and a visual asset that can be posted or tested. A generated campaign image or video may arrive faster; deciding whether it expresses the brand’s intent is still a separate job. Back in security, Anthropic is taking a different approach to widening access. The Verge reports that its new OSS Scanner offers open-source projects free, periodic vulnerability scans if they opt in. Anthropic says its strongest models will produce the reports. The important qualification: those reports are generated without human review or triage, so a flagged issue could be incorrect or invalid. A maintainer would still need to check it before treating it as a confirmed vulnerability. The scanner is one part of a broader cyber defense announcement. According to The Information, Anthropic also plans support for critical infrastructure operators, including threat research, on-site engineers and access to frontier Claude models. Those are announced offerings, not measured security outcomes. Together, the initiatives show two ways the company wants to put its models into defensive work: recurring scans for projects that opt in, and more direct support for infrastructure operators. In both cases, the value depends on what people do with the findings. That question of when an agent should defer to evidence leads into a paper posted Thursday. Imagine an agent retrieving information that contradicts what its model already believes. A correct-looking final answer alone does not tell you whether the agent noticed the conflict, worked through it or disclosed uncertainty. The researchers tested four agents in controlled and naturally occurring knowledge conflicts, with matched cases that had no conflict. They examined behavior throughout each task, not just the final answer. Higher task accuracy did not necessarily mean better handling of uncertainty: some configurations recognized a conflict along the way but failed to communicate unresolved uncertainty when their final answer was wrong. Agents also sometimes noticed conflicts early and lost track of them later. The caveat is that interventions which improved that behavior often reduced task accuracy in these tests. For teams evaluating agents, the takeaway is to inspect how an answer was reached and whether unresolved conflicts survive into the response—not just whether the agent usually gets the task right. ...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 Information reports that Mike Kubzansky, formerly chief executive of Omidyar Network, is launching Andaris, a network for large investors to share information about the AI industry. Its stated aim is to encourage more trustworthy practices. The AI data-center rush is also changing where infrastructure investors look for deals. The Information reports growing interest in assets such as landfills and rail terminals, as investors seek opportunities outside the most heavily pursued AI infrastructure projects. In chips, The Information reports that Nvidia plans to invest in d-Matrix, a developer of AI server chips that aims to compete with it. The investment is planned, not completed; the report frames it as part of Nvidia’s effort to make its technology compatible with rival chip designers. And TechCrunch reports that Goodfire has launched monitors that inspect an AI agent’s model while it works, calling on another AI selectively rather than having one read every action. Goodfire says that approach costs less, though no quantified savings are established here. 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 Monday with what's up next!