Overview
- OpenAI launched Daybreak on Tuesday, pairing GPT-5.5 model tiers with a Codex agent to review code, map attack paths, test exploits in isolation, and generate audit-ready patches.
- Daybreak runs across three tiers that include GPT-5.5-Cyber, a higher-capability variant with restricted access for trusted defenders under added identity checks and safeguards.
- The European Commission confirmed OpenAI will give EU cyber teams access to GPT-5.5-Cyber for evaluation and potential deployment as regulators assess high-risk AI for security use.
- Fresh evaluations by the UK AI Security Institute and Palo Alto Networks found Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 completing complex multi‑stage attack simulations and turning found bugs into working exploit paths faster than prior models.
- Google’s threat unit reported a criminal zero‑day likely created with AI, and the Pentagon is using Mythos to scan government systems even as it plans to phase out other Anthropic tools, reinforcing warnings that organizations have only three to five months to get ahead of AI‑driven exploits.