Latest / UpNext AI / Block’s Daybreak Security, Third Moment’s Agents, and Gemini 4 Argon
Transcript
- Welcome to the UpNext AI podcast. It's Thursday, October 15th, 2026, and here's what matters in AI today. Block is using OpenAI’s Daybreak cybersecurity platform to help its security team investigate alerts, test defenses, and propose fixes. OpenAI’s case study describes a workflow with human review built into the process: agents can investigate or suggest a patch, while people validate findings and guide deployment. The company says the system has increased Block’s triage throughput by 50 percent. In recent red-team exercises, the time from submitting a high-level vulnerability to an automated proposed patch was 18 minutes. That measures a proposed fix, not a deployed one. For another task, the team says it can replay a proposed detection change against seven months of production telemetry in under an hour, instead of manually revisiting old cases. The interesting operational detail is how Block handles context. Its security harness routes jobs to different models, lets humans take over, and can have agents check one another’s work. In one investigation described by Block’s security chief, an alert on a laptop led the model to connect evidence across endpoint and identity logs, exposing a broader set of issues. The team still checks the result before acting. For security leaders, this is a case study in shortening the path from alert to investigation and proposed remediation. The reported gains are Block’s experience, and its human approval step remains central when a suggested change could affect production systems. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast. That question of keeping agents useful over longer stretches leads to Third Moment Research. In an OpenAI case study, the lab describes running 10,000 agents at once on investigations lasting 36 to 48 hours. It uses models including GPT-6 Astra and GPT-6.1 Sol with Codex, and uses Daybreak to test the environments where those agents work. Its early work focuses on quantitative finance, where agents run experiments against historical data and decide what to test next. One problem was surprisingly mundane: a trading-strategy backtest took roughly three hours, leaving agents repeatedly checking whether it had finished. According to the case study, Codex identified caching and memory changes that cut that particular backtest to roughly three to five minutes. The immediate benefit was less waiting and fewer tokens spent on status checks. Third Moment also places an outer system around its working agents. When an agent leaves a task unfinished, that system decides whether to retry, stop, or send it for human review. The company is trying to improve agents by generating new test scenarios from recurring weaknesses. Its larger ambitions remain plans; the concrete example here is the combination of sustained investigations, faster supporting tools, and an escalation path when an agent gets stuck. For builders choosing a model for those workflows, price per token tells only part of the story. Developer Simon Willison has updated his llm-gemini plug-in to support Google’s Gemini 4 Argon, which became available through the Gemini API. His release note also points to access through Google’s monthly subscription plan. Willison lists the launch API price at $2 per million input tokens and $10 per million output tokens, with those prices set to double in December. He also cites an Artificial Analysis benchmark run that cost $2,407 in total with Argon. That is the cost of running that benchmark with the model, rather than a subscription price or a measure of quality by itself. His more useful buying question is how many tokens a model needs to finish a real task inside an agent workflow. The plug-in makes Argon easier to try, but teams should compare completed tasks and total run costs on their own work, especially with a price change scheduled. Willison also tried generating visual examples with the model; those demonstrations show a possible use, rather than a broad performance result. A research note on a different kind of capability: can a model read what someone believes and respond effectively in a conversation? A paper introducing the PerSo benchmark tests those skills together using persuasion dialogues from Reddit’s ChangeMyView community. The researchers built evaluation tasks with a model-assisted process and checked them through human annotation. The distinction is between understanding another person’s perspective and choosing a persuasive response. Across the models tested, performance generally improved with model size, but the two skills did not move together. On the response task, the strongest model slightly exceeded the best observed human performance. On the perspective-reading task, the best human participant remained ahead. That makes PerSo useful for teams evaluating conversational agents: a strong response score does not establish equally strong understanding of the other person. The result comes from this benchmark’s persuasion dialogues, so it should not be treated as a general measure of social intelligence. ...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. The Information reports that Anthropic has agreed to license technology from Opaque Systems and hire some of its engineers. Opaque sells software for securing confidential data used with AI agents, making this a move toward the data-security side of agent deployment. In Britain, WIRED reports that OpenAI and Anthropic told lawmakers voluntary AI-safety pledges are insufficient and called for new laws. Google DeepMind and Meta declined to back new legislation at the hearing. The hearing put a concrete policy split among major labs before Parliament; it did not enact new rules. And TechCrunch reports that Colossal Biosciences announced an AI platform intended to accelerate its de-extinction work. Peter Diamandis has launched a fund to back startups using the platform. The acceleration is the stated aim, rather than a demonstrated result. 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!