OpenAI’s New ROSALIND Is Now Performing At Human Level

OpenAI just dropped Rosalyn, a new AI model built for biology, drug discovery, and serious life sciences research. And that alone would be a big story. Though that was just the start because it also launched GPT 5.4 for cyber for defensive security work pushed new agents SDK upgrades for more secure AI agents and then the whole week took a darker turn with the shocking attack case tied to Sam Alman and OpenAI. There's a lot to get into here though the main headline is this new model. So let's start with that. So GPT Rosland is a purpose-built model focused on life sciences. We're talking biochemistry, genomics, protein engineering, drug discovery, and translational medicine. And the reason this matters is because the workflows in this field are insanely complex. You've got massive amounts of literature, databases, experimental results, evolving hypotheses, and everything is connected in ways that are hard to track even for experienced researchers. That's the exact problem OpenAI is going after here. Right now, if you look at drug development timelines, it typically takes around 10 to 15 years to go from initial target discovery to regulatory approval in the US. That's a huge bottleneck. And a lot of that time is spent in the early stages just figuring out what's even worth testing. OpenAI's angle here is that if you can speed up those early steps, the impact compounds across the entire pipeline. GPT Roslin is designed to work across those early stage workflows. It can synthesize evidence, generate hypotheses, plan experiments, and handle multi-step research tasks. That includes reading scientific papers, querying databases, using specialized tools, and even suggesting new experiments based on what it finds. So instead of just answering questions, it's acting more like a research assistant that can actually move a project forward. And the technical side is where things get more interesting. The model is optimized specifically for reasoning across molecules, proteins, genes, biological pathways, and disease related systems. It's also built to work with tools in a much more integrated way. So instead of just giving you text outputs, it can connect to scientific databases, run analyses, and help interpret results in context. They also introduced a life sciences research plug-in for codecs which connects the model to over 50 scientific tools and data sources that includes things like multiomix databases, literature repositories, protein structure tools, and more. The idea is to create a kind of orchestration layer where the model can move across different systems and combine insights in a way that's actually useful for real research workflows. And this isn't just theoretical. OpenAI is already working with some serious players in the space. Companies like Amgen, Madna, Thermo, Fiser Scientific, and the Allen Institute are using GPT Rosalin to apply it across real world workflows. There's also collaboration with Novo Nordisk to analyze complex data sets and identify potential drug candidates faster. The goal here is to explore more possibilities, find connections that would normally get missed, and generate better hypotheses earlier in the process. That's where the real leverage is. Now, performance-wise, OpenAI is showing some strong signals. On benchmarks like Bixbench, which focuses on real world bioinformatics and data analysis tasks, Rosland is outperforming other models with published scores. On labbench 2, which tests things like literature retrieval, sequence manipulation, and experimental design, it beats GPT 5.4 on 6 out of 11 tasks. And in specific tasks like molecular cloning design, the improvements are even more noticeable. They also ran evaluations with dynotherrapeutics using real unpublished RNA sequence data. In those tests, the model's best outputs ranked above the 95th percentile of human experts for prediction tasks and around the 84th percentile for sequence generation. That's not a small jump. That's already in the range where AI starts becoming a serious collaborator rather than just a tool. And there's a bigger context here. Since 2019, more than 17 billion has been invested into AIdriven drug discovery. Still, none of these AI developed drugs have reached large-scale trials yet. So, the industry is clearly still early. Everyone is pushing, though no one has fully cracked it. That's why moves like this matter. Open AAI is trying to position itself right in the middle of that transition point. At the same time, they're being very careful about how this gets deployed. Rosalind is not fully open. It's released as a research preview through a trusted access program. That means only qualified organizations can use it, and there are strict controls around access, governance, and safety. The idea is to make sure it's used for legitimate scientific work with proper oversight and security in place. They're also putting a lot of emphasis on enterprisegrade controls, access management, compliance requirements, internal governance. All of that is part of the roll out. Organizations have to prove beneficial use, maintain proper safeguards, and restrict access to approved users. So yeah, this is a very different kind of launch compared to something like GPT4 or GPT5. It's not about mass adoption right away. It's about building a foundation inside a highstakes domain where the impact could be massive if it works and OpenAI is clearly planning to expand this. Rosalind is described as the first model in a broader life sciences series. They're already talking about pushing further into long horizon workflows, deeper biochemical reasoning, and more advanced tool integration. There's even mention of ongoing work with institutions like Los Alamos National Laboratory on things like protein and catalyst design. So, this is not a one-off. This is the beginning of a long-term push into scientific discovery. At the same time, while they're moving into biology, they're also doubling down on cyber security, which brings us to GPT5.4 cyber. This is another specialized version of their flagship model, though focused entirely on defensive security work. And the key difference here is that some of the usual restrictions have been relaxed specifically for legitimate security professionals. Normally models have strict limits to prevent misuse. That makes sense for general users, though it can get in the way when you're actually trying to analyze vulnerabilities or investigate threat. GPT 5.4 Cyber is designed to remove some of those barriers so professionals can do their work more effectively. One of the standout features is binary code analysis. So instead of needing source code, the model can analyze compiled software directly. That means it can look for vulnerabilities, detect potential malware behavior, and assess security risks even when the underlying code isn't available. It also supports extended workflows for things like vulnerability research, defensive programming, and security training. So again, similar pattern to Rosalind. It's not just about answering questions. It's about handling multi-step processes in a more integrated way. Access here is also restricted. It's part of OpenAI's trusted access for cyber program where individuals and organizations need to be verified before they can use the model. The idea is to scale access responsibly while still maintaining control over how these capabilities are used. And this is happening in direct response to what Anthropic is doing with Claude Mythos preview. That model is reportedly capable of finding and exploiting vulnerabilities in a highly autonomous way, though it hasn't been released publicly because of the risks. Instead, Anthropic launched Project Glasswing with a small group of partners like AWS, Apple, Google, Microsoft, and Crowdstrike. OpenAI is taking a slightly different approach. Rather than limiting access to a small group, they're building a broader system with identity verification and tiered access. The goal is to eventually support thousands of individuals and hundreds of teams while still keeping safeguards in place. They're also emphasizing three core principles for their cyber strategy. First, democratized access, meaning legitimate defenders should have access to advanced tools. Second, iterative deployment where models are rolled out gradually with continuous improvements. And third, investment in ecosystem resilience, which includes things like grants, open- source initiatives, and tools like Codeex Security. Speaking of Codex Security, that system has already contributed to fixing more than 3,000 critical and high severity vulnerabilities. And through the Codeex for open-source program, over 1,000 open-source projects have received free security scans. So, the direction here is clear. AI is becoming a major factor in cyber security and both OpenAI and Anthropic are racing to build systems that can operate at a much higher level than traditional tools. Then there's another piece of the puzzle that connects to both Rosland and Cyber, and that's OpenAI's agents SDK update. This one is more focused on developers and enterprise use, though it still plays into the bigger picture. The update introduces a model native harness that allows agents to work across files and tools directly on a computer. There's also a new sandbox environment for execution along with configurable memory and orchestration systems. Basically, it makes it easier to build and deploy AI agents that can actually do things, not just generate output. Before this, developers had to manage a lot of the infrastructure themselves. Now, OpenAI is trying to simplify that by providing more built-in capabilities. So you can build an agent inside their ecosystem and rely on their tools for things like memory management, execution environments, and security. From a business perspective, this is also about keeping developers inside the OpenAI ecosystem. The easier it is to build agents using their tools, the more likely developers are to stay within that environment, and that increases usage, token consumption, and overall demand for their services. There's a trade-off, though. Some companies prefer to remain provider agnostic and build their own systems. For them, this kind of tightly integrated approach might be less appealing. So, it's a balance between convenience and control. Still, the trend is obvious. Open AAI is pushing toward a world where AI agents can operate across complex workflows, whether that's scientific research, cyber security, or enterprise applications. And then there's one more story that shows how intense things are getting around AI right now, and that's the situation involving Sam Alman. A man in Texas has been charged with attempted murder after allegedly attacking Altman's home in San Francisco with a Molotov cocktail. According to the US Justice Department, the suspect also attempted to target OpenAI's headquarters shortly after. He was found with documents criticizing AI and calling for violence against executives and investors in the industry. The charges include attempted murder, possession of an unregistered firearm, and destruction of property using explosives. Authorities say he set fire to an exterior gate at Altman's home and later tried to break into the open AI building using a chair. Investigators also recovered incendiary devices, kerosene, and a lighter. No one was injured, though the situation clearly highlights how heated the conversation around AI has become. Alman himself responded by saying, "There needs to be a shift toward less aggressive rhetoric and more constructive discussion." At the same time, this incident came just hours after a critical profile of Altman was published, questioning his leadership and trustworthiness. So, the timing added another layer to an already tense situation. And zooming out, this kind of reaction, even though it's extreme, shows how much impact AI is starting to have. It's not just a tech story anymore. It's affecting public perception, policy debates, and even personal safety. And one more thing, our science channel, Science Revolution, already has over 700 subscribers. If you're into science and space, go check it out. Link in the description. Subscribe, share it with a friend. We'd really appreciate it. Anyway, that's it for this one. Let me know what you think in the comments. And if you enjoyed the video, drop a like and subscribe. Thanks for watching, and I'll catch you in the next one.