Latest / UpNext AI / OpenAI’s Astra Cyber Model, AfterQuery’s $3.2 Billion Valuation, and Multi-Day Coding Agents | UpNext AI – September 2, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Wednesday, September 2nd, 2026, and here's what matters in AI today. OpenAI says it will soon release Astra, a large language model whose most advanced cybersecurity capabilities will be more tightly limited. According to the company, Astra is its first model to meet OpenAI’s “critical cybersecurity threshold.” That designation matters because OpenAI says the model can find previously unknown security flaws and exploit them without a person guiding the process. On ExploitBench, which tests an LLM’s ability to hack known vulnerabilities, OpenAI says Astra achieved a perfect score. The company also says that, in a modified internal test, it found and exploited two zero-day vulnerabilities. The rollout plan is as important as the capability claim. OpenAI says it will first preview Astra with a group of testers, while restricting access to its most advanced cyber features. It has also described work on improved abuse detection, jailbreak prevention, higher-risk account controls, and chain-of-thought monitoring intended to spot harmful behavior. There is still plenty we do not know: who will test Astra, how those testers will be selected, and what the full external evaluation record will show. OpenAI says it plans to release additional evaluations and safety information when the model launches more broadly. For security teams, the near-term message is simple: frontier-model releases are increasingly arriving with both new defensive tools and a more credible offensive capability attached. That collision between capability and control is one side of the AI market. The other is the scramble to supply the expertise that trains these systems. TechCrunch reports that AfterQuery, an AI model-training startup, has raised a round at a reported $3.2 billion valuation. That is a dramatic jump from the $30 million Series A it announced in April, when it was valued at $300 million—just five months ago. According to Y Combinator partner Gustaf Alströmer, the company is the accelerator’s fastest startup to reach unicorn status. AfterQuery says it works with knowledge professionals, including doctors, lawyers, and other specialists, to train models and agents to perform professional tasks—not merely answer questions correctly. The company had said in April that it reached a $100 million annualized revenue run rate and named Nvidia, Legora, and Korea’s Motif Technologies among its customers. Forbes first reported the new round, and AfterQuery was not immediately available for comment. A valuation is not a verdict on a business, but this one is a sharp signal. Investors are placing enormous value on the human expertise, task data, and operational know-how needed to make AI agents useful in real work. For the research note, a new arXiv preprint tackles a familiar frustration with coding agents: a model can produce a promising first pass, then lose the plot as a project grows. The paper proposes Harness-of-Harness, or HoH, a control layer that sits on top of existing coding-agent setups. Rather than asking an agent to build everything in one sweep, it organizes repeated planning, coding, and testing loops. It also breaks work into smaller verifiable increments, separates implementation testing from independent evaluation, and keeps versioned project histories. The researchers tested HoH with three model-and-harness pairs across GameCraft-Bench, FrontierSWE, and ProgramBench. They report that the layered setup outperformed the corresponding standalone harnesses by an average relative gain of 52.25 percent, with a maximum gain of 82.86 percent after three iterations. In a separate multi-day demonstration spanning more than 70 iterations, the system built a first-person-shooter game with a storyline, core mechanics, visuals, and audio. That is an intriguing proof of persistence, not proof that autonomous engineering is ready to replace a software team. This is an arXiv preprint, and its results are benchmark and demonstration results rather than evidence from deployed engineering organizations. Still, the direction is clear: the harness around a coding model may matter as much as the model inside it. ...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. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ... Anthropic has launched Claude Fable 5.1 and Mythos 5.1. The company says Fable 5.1 improves coding and research work while cutting the cost of agentic tasks by up to 45 percent. For teams running agents at scale, lower token costs can matter as much as a benchmark win. Google has published a roundup of the AI updates it announced during August. It is a consolidation rather than a new product launch, but useful for anyone trying to catch up on the company’s latest AI releases without chasing a month of announcements. The Financial Times reports that GoPro is set to be acquired by an AI hardware group in a $285 million cash deal. GoPro shares rose as much as 80 percent after the announcement, a reminder that the AI hardware land grab now reaches well beyond chips and data centers. 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. If you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!