AI Daily — June 23, 2026
2026-06-23
openai.com
OpenAI Daybreak: Codex Security and GPT-5.5-Cyber for Vulnerability Management
OpenAI launched Daybreak, a security-focused initiative introducing Codex Security and a new GPT-5.5-Cyber model designed to find, validate, and patch vulnerabilities at scale across organizations. The tooling targets the full vulnerability lifecycle — detection through remediation — rather than just flagging issues. This marks OpenAI's first explicit push into enterprise security infrastructure with purpose-built models.
openai.com
Patch the Planet: OpenAI's AI-Assisted Open Source Vulnerability Remediation
Alongside Daybreak, OpenAI announced Patch the Planet, a program pairing AI tooling with expert human review to help open-source maintainers identify and fix security vulnerabilities. The initiative addresses the under-resourced reality of OSS security maintenance by applying Codex-based automation to triage and patch generation. This is a meaningful applied use of code-capable LLMs in a high-stakes, underserved domain.
www.technologyreview.com
Anthropic's Feud with the US Government Over 'Mythos' Model
Anthropic is in a dispute with the US government stemming from the company's April disclosure of an internal model called Mythos, raising regulatory and oversight tensions. MIT Technology Review outlines three key dynamics to watch, including questions about voluntary disclosure obligations and government access to frontier model evaluations. The situation highlights unresolved friction between frontier AI labs and federal oversight frameworks.
jack-clark.net
AI Decisively Outperforms Expert Humans at Persuasion — Oxford/Stanford/AISI Study
A multi-institution study from Oxford, Stanford, and the UK AI Security Institute found that AI systems were reliably more persuasive than expert human communicators across tested conditions — a capability the researchers term 'superpersuasion.' The finding has direct implications for influence operations, misinformation, and the tractability of AI-mediated manipulation at scale. Jack Clark's Import AI 462 covers this alongside research on self-sustaining AI systems and paths to ASI.