Image via Unknown
Tuesday, December 23, 2025
OpenAI's GPT-5.2-Codex is here to secure your code
OpenAI dropped GPT-5.2-Codex with beefed-up cybersecurity chops for coding, while Google DeepMind countered with FunctionGemma, an open-source model that lets you run AI agents offline and privately (bold move for the sovereignty crowd). Meanwhile, Anthropic's having a moment with Claude now available in Chrome for all paid users AND their vending machine experiments showing that even smart AI can be hilariously easy to fool (yikes). Oh, and Greptile's data shows AI coding tools are boosting developer velocity by 76% with Anthropic closing the gap on OpenAI fast, so the competitive heat is very much on. If guardrails keep breaking this easily, how safe are we really?
Image via Unknown
Top Stories
OpenAI
OpenAI launched GPT-5.2-Codex with frontier agentic coding and strengthened cybersecurity capabilities, already proven to discover real vulnerabilities in production software. The company is deploying it through a staged approach with safeguards and exclusive access programs to maximize defensive security impact while mitigating dual-use risks.
Anthropic
Claude-powered AI agents running real vending machines showed dramatic operational improvements in Phase Two but remained surprisingly vulnerable to social engineering and poor judgment, revealing that deploying autonomous AI in the real world requires far more sophisticated guardrails than current approaches provide.
Google DeepMind
Google DeepMind's FunctionGemma enables lightweight, on-device AI agents that execute functions from natural language commands with full privacy and minimal latency, positioning small specialized models as the foundation for practical edge AI systems.
Greptile
AI coding tools are dramatically boosting developer productivity, with PR sizes up 33% and code output per developer up 76%, while Anthropic's rapid SDK adoption narrows OpenAI's market dominance in the AI coding ecosystem.
Anthropic
Anthropic releases Claude in Chrome extension for all paid users, enabling AI-powered browser automation for research, data extraction, and task completion; the feature integrates with Cowork for seamless workflow automation but requires careful security practices.
Keep Reading
Industry Voices
Shubho Sengupta
Co-founder at Axiom Math
Building tools that turn math into verifiable code, making AI reasoning actually checkable instead of just plausible.
RJ Marsan
Engineer at OpenAI
Ships the infrastructure behind ChatGPT's actual deployment, so you get the engineering reality behind the hype.
Enjoyed this issue?
Get daily AI intel delivered to your inbox. No fluff, just the stories that matter.