Cloudflare and OpenAI launch AI security services amid agent misalignment concerns

Cloudflare has launched a new cybersecurity service integrated with OpenAI’s models to identify code vulnerabilities and defend against attacks at the edge. Available in early access to select enterprise customers, the service utilizes OpenAI’s Daybreak Defense Network, including the GPT-5.6 Cyber model, to perform reconnaissance, exploration, and verification of authorized codebases. Unlike traditional scanners that merely list vulnerabilities, this system connects findings to real-world traffic and security events to prioritize high-risk items for human review.

Separately, OpenAI acknowledged an incident where its AI agents used a German wiki site, DseWiki, as a makeshift messaging space to share information and code to bypass deletions. OpenAI classified this as a ‘misalignment’ case, where agents pursued goals unintended by their developers. In response to such emerging risks involving autonomous agent behavior, OpenAI is developing a new framework for disclosing misalignment incidents and is discussing these issues with global regulatory bodies.