Today's AI news: OpenAI finds PHP zero-days, launches Decisions API, Uber deploys ChatGPT to 30k staff, and DeepMind maps global forests.
Here is what happened in AI today, Thursday 23 October 2026 — the biggest stories from across the industry.
OpenAI's Security Program Discovers Critical PHP Vulnerabilities Affecting 70% of the Web
In one of the most consequential AI-assisted security disclosures in recent memory, OpenAI's Patch the Planet program — using GPT-5.6 Sol — was one of only two parties to independently discover a set of PHP2Shell vulnerabilities in core PHP. Because PHP underpins roughly 70% of the web, the exposure was enormous. OpenAI and the co-discoverer coordinated a responsible disclosure process that ran for two months before going public today. The episode is being described internally as "The Big Hardening" and signals a new era in which frontier AI models are actively used as offensive security research tools to harden critical internet infrastructure before adversaries can exploit the same flaws.
OpenAI Launches the Decisions API
OpenAI has released a new Decisions API designed for developers who need fast, focused, and low-cost binary or categorical decisions inside their applications. Rather than routing every query through a large, general-purpose model, the Decisions API is purpose-built for high-throughput decision logic — think content moderation flags, routing rules, or eligibility checks — without the latency or cost overhead of a full conversational model. The launch reflects OpenAI's continuing push to offer tiered, task-specific inference products alongside its flagship models.
Uber Rolls Out ChatGPT Work to More Than 30,000 Employees
Uber has struck a company-wide deal with OpenAI to deploy ChatGPT Work across its entire global workforce of more than 30,000 employees. Staff will use the platform for research, data analysis, presentation creation, and complex multi-step workflows. The rollout is one of the largest enterprise ChatGPT deployments announced to date and underscores how quickly AI productivity tools are moving from pilot programmes to standard corporate infrastructure. OpenAI's forward-deployed engineering (FDE) model — which embeds OpenAI engineers directly with enterprise customers — is understood to have supported the integration.
AI Discovers a New Consensus Protocol — Beating State-of-the-Art by Up to 33%
Researchers have published details of A-greement, a new consensus protocol for geo-replicated state machines discovered with the help of an AI agent harness called AI-greement. Averaged across six deployments, A-greement reduces mean latency in a geo-distributed key-value store by 12.2% on write-heavy workloads and 33.3% on read-heavy workloads compared to the fastest existing baseline. The AI-greement harness integrates a protocol knowledge base, a modular implementation skeleton, and a Lean-based formal verification framework — and along the way uncovered previously unreported vulnerabilities in existing consensus protocols. The authors are careful to note that human experts remained essential throughout the process, arguing that AI reshapes the researcher's role rather than replacing it.
Google DeepMind Maps Every Forest Type on Earth at 10-Metre Resolution
Google DeepMind has released a global forest-type map that categorises every 10-metre patch of land on Earth into distinct forest classes. The project goes well beyond simple tree-cover detection — it distinguishes between different forest types at a granularity that was previously impractical at planetary scale. The map has immediate applications in biodiversity monitoring, carbon accounting, and climate policy, and represents a significant milestone in the use of AI for large-scale environmental intelligence.
NVIDIA Kumo Outperforms Supervised ML on Relational Data — No Training Required
NVIDIA has announced that its Kumo Relational system outperforms traditional supervised machine learning on relational data tasks without requiring any task-specific training. The result is significant for enterprises that hold large relational databases — customer records, transaction logs, inventory tables — but lack the labelled datasets or ML engineering resources to build bespoke models. Kumo's zero-training approach lowers the barrier to production-grade ML on structured data considerably.
That's the main news for today. Follow ItsTimeAi for tomorrow's roundup at itstimeai.com.au.
Frequently Asked Questions
What did OpenAI's Patch the Planet program discover?
OpenAI's Patch the Planet program, using GPT-5.6 Sol, was one of two parties to independently discover PHP2Shell vulnerabilities in core PHP. Because PHP runs on roughly 70% of the web, the vulnerabilities were critical. A coordinated two-month disclosure process preceded today's public announcement.
What is OpenAI's new Decisions API?
The Decisions API is a fast, focused, and low-cost OpenAI product designed for high-throughput decision logic inside applications — such as content moderation, routing rules, or eligibility checks — without the overhead of a full conversational model.
How many Uber employees are getting access to ChatGPT Work?
Uber is rolling out ChatGPT Work to more than 30,000 employees globally, making it one of the largest enterprise ChatGPT deployments announced to date. Staff will use it for research, data analysis, presentations, and complex workflows.
What is the A-greement consensus protocol and how was AI involved in finding it?
A-greement is a new consensus protocol for geo-replicated state machines that reduces latency by up to 33.3% over the fastest existing baseline. It was discovered using an AI agent harness called AI-greement, which combined a protocol knowledge base, modular implementation tools, and a formal verification framework. Human experts remained central to the process throughout.