OpenAI said on March 9, 2026 that it plans to acquire Promptfoo. The company said Promptfoo's technology will strengthen agentic security testing and evaluation inside OpenAI Frontier, while Promptfoo remains open source under its current license and existing customers continue to receive support.
OpenAI says Codex Security deliberately does not start from a SAST report because many real vulnerabilities come from broken validation order, canonicalization, and other behavioral flaws rather than simple dataflow patterns. Instead, the system starts from repository behavior and validates hypotheses with focused tests in a sandbox.
On March 5, 2026, OpenAI introduced GPT-5.4 as a flagship model focused on relevance, contextual understanding, and instruction following. In the API, it pairs a 1M-token context window with stronger tool search for long, multi-tool workflows.
OpenAIDevs said on March 16, 2026 that subagents are now available in Codex. The feature lets developers keep the main context clean, split work across specialized agents, and steer individual threads as they run, while the official docs already describe PR review and CSV batch fan-out patterns.
OpenAI said on March 10, 2026 that its new IH-Challenge dataset improves instruction hierarchy behavior in frontier LLMs, with gains in safety steerability and prompt-injection robustness. The company also released the dataset publicly on Hugging Face to support further research.
OpenAI said on February 20, 2026 that its theorem-proving model produced proof attempts for all 10 research-level First Proof problems. After expert feedback, the company believes at least five attempts are likely correct, while some remain under review and the attempt for problem 2 now appears incorrect.
On March 11, 2026, OpenAI published new guidance on designing AI agents to resist prompt injection, framing untrusted emails, web pages, and other inputs as a core security boundary. The company says robust agents separate data from instructions, minimize privileges, and require monitoring and user confirmation before taking consequential actions.
On February 5, 2026, OpenAI launched Trusted Access for Cyber, an identity-verified program that gives defenders controlled access to enhanced cyber capabilities built on GPT-5.3-Codex. OpenAI paired the pilot with automated monitoring and a $10 million grant program for security researchers and infrastructure defenders.
On March 9, 2026, OpenAI said it plans to acquire Promptfoo and integrate its AI security tooling into OpenAI Frontier. The move pushes security testing, red-teaming, and governance closer to the default workflow for enterprise agents.
On March 6, 2026, OpenAI reposted a message from Sachin Katti saying construction is underway in Port Washington, Wisconsin. The post turns OpenAI’s previously announced Stargate and partner-led compute strategy into a visible on-the-ground build milestone.
OpenAI said on March 5, 2026 that GPT-5.4 Thinking shows low Chain-of-Thought controllability, which for now strengthens CoT monitoring as a safety signal. The release pairs an X post with a new open-source evaluation suite and research paper.
OpenAI introduced ChatGPT for Excel in beta on March 5, 2026 and added new financial data integrations inside ChatGPT. The launch targets spreadsheet modeling, due diligence, and research workflows built on GPT-5.4.