OpenAI has released GPT-5.6-Cyber, a specialized model designed to help cybersecurity teams identify vulnerabilities before attackers exploit them. The system answers 98.5 percent of security queries that would normally be restricted.
OpenAI's new GPT-5.6-Cyber model represents a shift in how AI assists cybersecurity defense. The system successfully handles security-related queries at a 98.5 percent success rate, providing defenders with detailed vulnerability analysis that previous models would have blocked.
The model has already demonstrated practical value. During testing, GPT-5.6-Cyber uncovered two previously unknown vulnerabilities in Chrome, validating its ability to surface genuine security flaws.
OpenAI frames the release as a response to shrinking response windows in cybersecurity. As attack sophistication increases, the time available for defenders to patch vulnerabilities before exploitation narrows. By providing AI-assisted vulnerability discovery, the company argues defenders can work faster than attackers.
The tool targets security professionals and organizations conducting authorized penetration testing and vulnerability research. Access requires identity verification to prevent misuse by bad actors.
The 98.5 percent query success rate marks a significant improvement over previous model restrictions. Earlier versions of GPT systems included safeguards that blocked many security-related questions entirely, limiting their usefulness for legitimate defense work.
GPT-5.6-Cyber's release reflects broader industry debate around AI and security. Stakeholders have long argued that restricted AI tools disadvantage defenders relative to attackers, who operate without such constraints. Others raise concerns about democratizing vulnerability discovery tools.
OpenAI's identity verification requirement attempts to balance these competing interests—enabling legitimate security work while adding friction for potential misuse.
The model joins other AI tools entering the cybersecurity space. Security firms have increasingly integrated large language models into their offerings, automating threat analysis and incident response.
GPT-5.6-Cyber's Chrome vulnerability discoveries suggest the tool can identify genuine security issues. Whether it scales to enterprise vulnerability discovery programs remains unclear, as does long-term accuracy and false positive rates in real-world deployment.
Kinney Drugs has rolled back its AI-powered phone assistant after receiving hundreds of customer complaints. The pharmacy chain cited service issues as the reason for the withdrawal.
A collaborative project from Hugging Face and EleutherAI tested 14 open-source OCR models on historical texts, finding that the best performer achieves 97.6% character accuracy at under $2 per thousand pages—sufficient for training language models.
Meta released Muse Glimmer, a downloadable AI model that runs on personal computers with a single GPU. The move intensifies Meta's push for accessible, open-source AI against closed competitors.
EY is hiring a head of "agent economics" to manage its expanding AI-powered workforce, signaling that major corporations are scrutinizing whether expensive AI investments deliver returns.