In July 2026, researchers from HacktronAI executed a sophisticated attack against OpenAI by chaining a heap buffer overflow in the libheif library with an SSO misconfiguration. The exploit targeted the Discourse-hosted community forum (community.openai.com), where an image-upload pipeline passed HEIC/HEIF files to ImageMagick. Due to the forum's Debian 12 base, it utilized a vulnerable libheif version (1.19.7) that lacked critical security backports, allowing for out-of-bounds read/write primitives. Using Anthropic's Claude Opus 5 models to automate exploit development and porting from ARM64 to x86-64/jemalloc environments, the team achieved remote code execution (RCE) within hours. This forum compromise allowed for the takeover of ChatGPT and Codex accounts via OpenAI's identity infrastructure. To demonstrate impact, the researchers used an employee's Codex account to open pull request #1186742 in OpenAI's internal monorepo, openai/openai. The entire operation, from initial discovery to repository access, was completed in under 72 hours. OpenAI subsequently paid a $6,500 bounty for the disclosure. The researchers noted that AI models significantly compressed the timeline for operationalizing memory corruption vulnerabilities, effectively turning scarce human expertise into scalable compute.