OpenAI's specialised models
OpenAI rolled out GPT-Rosalind, a model aimed at life‑sciences research with stronger biology knowledge, and a cybersecurity-focused GPT-5.4-Cyber that eases restrictions for verified professionals but will be tightly limited in access. ( ). At the same time Europe’s AI rules are raising practical compliance questions — the EU AI Act adds specific logging requirements for AI agents while research flags that its human-rights safeguards largely stop at the bloc’s borders. ( ).
OpenAI is splitting its latest artificial intelligence work into narrower tools: one for biology labs, one for cyber defenders. (openai.com) On April 16, OpenAI introduced GPT-Rosalind, a model built for biology, drug discovery and translational medicine, and said the new life-sciences series is tuned for chemistry, protein engineering and genomics workflows. OpenAI said the model is available as a research preview through ChatGPT, Codex and the application programming interface under its trusted-access program. (openai.com) Two days earlier, on April 14, OpenAI said it was expanding its Trusted Access for Cyber program and adding GPT-5.4-Cyber, a version of its flagship model tuned for defensive cybersecurity work. The company said access would go first to thousands of verified individual defenders and hundreds of teams protecting critical software. (openai.com) The split reflects the kind of work these fields do. Drug research starts with targets, molecules and lab evidence that can take 10 to 15 years to turn into an approved medicine, while cyber defense revolves around finding software flaws before attackers do. (openai.com, euronews.com) OpenAI is also loosening one set of safeguards while tightening who gets in. Euronews reported GPT-5.4-Cyber is more permissive on requests tied to vulnerability discovery, but the rollout is tightly limited to vetted professionals rather than the public. (euronews.com, openai.com) Europe’s rules are moving in the opposite direction: toward more documentation around how artificial intelligence systems behave in the field. Article 12 of the European Union AI Act requires high-risk systems to support automatic event logging over their lifetime so operators and regulators can trace decisions, monitor failures and investigate risks. (ai-act-service-desk.ec.europa.eu, helpnetsecurity.com) Those logging duties do not attach to every model by default. Help Net Security noted that a general-purpose model becomes subject to high-risk system obligations when another company integrates it into uses such as healthcare, employment, finance or law enforcement. (helpnetsecurity.com, theup.io) Another gap sits outside Europe’s market. A research summary published by Global Voices said the AI Act does not stop companies from exporting systems that would be banned or treated as high-risk inside the European Union, and the authors called for human-rights due diligence to follow those exports. (globalvoices.org) That leaves companies building specialized models with two separate tests in 2026: prove the tool is useful enough for scientists or security teams to adopt, and prove the system around it can survive tighter access controls, logs and audits. (openai.com, openai.com, ai-act-service-desk.ec.europa.eu)
