OpenAI announces the expansion of its Trusted Access for Cybersecurity (TAC) initiative and launches the customized large model GPT-5.4-Cyber. This model lowers the rejection threshold for security defense tasks, granting verified security teams access to binary reverse engineering capabilities without requiring source code. This move directly pushes the application of large models into foundational system defense and code analysis. In February 2026, OpenAI initiated the TAC project and established a $10 million cybersecurity grant program. On March 6, Codex Security entered the research preview phase, covering over 1,000 open-source projects. As of April 14, the system has helped fix over 3,000 high-risk and critical vulnerabilities. For access, enterprises must apply through OpenAI representatives, while individual users can complete KYC identity verification via http://(chatgpt.com)/cyber to obtain tiered model permissions based on verification signals. Traditional Static Application Security Testing (SAST) tools, plagued by high false positive rates, have become casualties in this evolving ecosystem. Codex Security bypasses static analysis reports in its architecture, opting instead to call z3-solver in Python environments to handle input constraint problems. It extracts code slices by writing micro fuzzing programs and directly generates end-to-end proof-of-concept (PoC) for vulnerabilities. Traditional cybersecurity vendors, reliant on periodic audits and static vulnerability lists, are seeing their service models squeezed by AI agents automating risk reduction processes. External industry actions were a direct catalyst for this product release. On April 7, competitor Anthropic launched Project Glasswing and its dedicated cybersecurity model Mythos, granting testing access to 40 organizations including Apple, Google, and Microsoft. They also disclosed a 27-year-old hidden vulnerability in OpenBSD captured using their model. A week later, OpenAI followed up with the release of GPT-5.4-Cyber. NETGEAR’s Product Security Lead, Chandan Nandakumaraiah, commented that the analysis results provided by OpenAI “often feel like having a seasoned product security expert working alongside you.”