OpenAI halts model training: Nvidia says it can make working with agents safer

Gemini-3.8 Flash TTS and Flash-Lite TTS speech models. Using AI models to speed up site/app UX. Cyber Index Alliance supports security benchmarks. Scientist says Anthropic plagiarized his team’s data.

Share
A team of AI engineers monitors a swarm of agents

In today’s edition of Data Points, you’ll learn about our top headlines, and more:

  • Gemini-3.8 Flash TTS and Flash-Lite TTS speech models
  • Using AI models to speed up site/app UX
  • Cyber Index Alliance supports security benchmarks
  • Scientist says Anthropic plagiarized his team’s data

But first:

OpenAI shuts down model training after security incidents

OpenAI paused training of its latest AI models after disclosing incidents in which its agents, while searching federal government websites, acted beyond what they were asked to do. In one case involving the Department of Education, agents found API developer keys to access government data, even though only publicly available information was gathered. In a separate case involving the SEC, agents took information that was freely available and posted it elsewhere online. AI evaluator Transluce separately reported that agents appearing to come from OpenAI tried and failed to hack into a Department of Education website, a claim OpenAI has not confirmed. OpenAI says it will resume training only once it has added safeguards, and expects to pause development again if similar issues arise. (Associated Press)

Nvidia releases tools to rein in wayward AI agents

Nvidia released the Open Agent Safety Platform, software meant to help developers test and monitor AI agents and stop them from acting outside their intended boundaries. Nvidia says the platform includes (i) a watchdog that can quarantine misbehaving agents within milliseconds and (ii) tools to control what agents can see, access, and interact with. The release follows several incidents in which AI agents from OpenAI, Anthropic, Meta, and Google went rogue, including an OpenAI agent that Australian Prime Minister Anthony Albanese said infiltrated a government-services website, and Anthropic’s disclosure that its test software hacked outside companies without its knowledge in three cases since April. Nvidia said each incident followed the same pattern: agents escaped sandbox safety controls to complete their assigned tasks. For developers building AI agents, the platform offers a vendor-backed option for enforcing boundaries and monitoring behavior in production; however, there is currently no independent verification of how well it prevents the sandbox-escape pattern Nvidia describes. (The Wall Street Journal)

Google’s new speech models let you design your own voices

Google released two new text-to-speech models, Gemini 3.8 Flash TTS and Flash-Lite TTS, available now through the Gemini API and Google AI Studio. Flash TTS targets creative direction: It lets users design original voices from natural-language prompts, access more than 2,000 production-ready voices across 100-plus languages, and clone a voice from a 30-second sample given required verification of consent. Flash-Lite TTS is built for high-volume, lower-cost use cases like dubbing and voice agents. Both models support long-form audio generation with minimal speaker drift, two-speaker scene staging, and non-verbal cues like laughs or sighs for scripted delivery. Google says Flash TTS ranks first overall on Hume AI’s Voice Design Benchmark (71.4) and both models place in the top two on Hume’s Overall Quality Index. (Google)

Anthropic uses Claude to speed up web and app loading times

Anthropic says an internal model made the core UX of Claude.ai and the Claude desktop app about three times faster during a two-week sprint in August. The model, run through a beta tool called Claude Tag, worked through a Slack channel with standing instructions to monitor deployments, build benchmarks, and propose fixes. Over the two weeks, it merged over 3,000 changes with human review and no customer-facing rollbacks. Anthropic reports that load time for a page on claude.ai dropped from 3.1 seconds to 0.55 seconds at the 75th percentile, and Claude Code session start time fell from 0.8 to 0.3 seconds. Some of the techniques the team used include instruction-count profiling with a tool called Valgrind, a static HTML composer that lets users type before React loads. Anthropic says the approach shows that giving a coding model fast, reliable ways to measure a problem, rather than just asking it to fix bugs, can let it iterate on performance work largely unsupervised as long as humans keep gating deployments and reviewing pull requests. (Anthropic)

Benchmark index measures models’ abilities to find and patch security flaws

Artificial Analysis launched the Cyber Index Alliance, a group including Collinear AI, IBM, Nvidia, and Vercel. The independent testing firm also released a new Cyber Index benchmark that rates AI models on enterprise cyber defense work. The index combines three evaluations: CWE-Bench-AA (120 tasks auditing real repositories against OWASP Top 10 vulnerability categories), DeepsecBench-AA (scoring vulnerability discovery against expert-verified findings), and CyberGym-E2E-AA (131 tasks requiring models to discover, reproduce, and patch memory-safety bugs in C/C++ projects like FFmpeg and CPython). Models work only from source code and are never asked to build working exploits, and all three evaluations run on Stirrup, the company’s open-source agent harness. In testing, 55% of CWE-Bench-AA failures were partial fixes that left related vulnerabilities open, over-corrections that broke legitimate functionality were more common in top-scoring models, and the best model on DeepsecBench-AA caught only 41% of expert-verified issues. The benchmark gives security teams a way to compare models’ abilities to discover vulnerabilities and patch holes specifically, rather than their general coding ability. (Artificial Analysis)

Scientist accuses Anthropic of plagiarizing team’s biology work

Anthropic said its AI agents used Claude to independently discover new enzymes it calls ARTs (array-associated reverse transcriptases) by searching gene databases. But Copenhagen computational biologist Mario Rodríguez Mestre says he and colleagues had studied the same enzymes, which they nicknamed ‘jumbotrons,’ for four years. The team shared unpublished findings with Claude while using it for research tasks. Mestre says a key finding Anthropic credited to its AI, involving RNA molecules tied to the enzymes, matches work he says his team completed over a year earlier and had already discussed with Claude. Anthropic responded that it found no prior published work on the ART system and that Claude was not trained on user transcripts. The dispute echoes a similar case this month in which OpenAI claimed to have solved a math problem that mathematician Tristan Buckmaster said he had been working on using OpenAI’s models. Mestre says he is now winding down his use of Claude and moving to other AI models over concerns his research may have been used without attribution. (The New York Times)


Want to know more about what matters in AI right now?

Read the latest issue of The Batch for in-depth analysis of news and research.

Last week, Andrew wrote about the difference between 0 to 1 prototypes and mature projects with many users. Each requires its own tactics for shaping the build and deploying architecture — two the core nodes in our AI Engineering Skills map. 

“It is not helpful to over-design at the early stages, or under-design a mature product. For an early stage project, if your primary goal is to quickly test a product idea, a casual software architecture design might be okay, with minimal thought given to efficiency, data schemas, cost of third party services, and so on. But as a project matures, giving careful thought to tradeoffs like latency, availability, consistency, reliability, maintainability, simplicity, and cost will give you a better outcome.”

Read Andrew’s letter here.

Other top AI news and research we covered in depth:


A special event for our community

AI Dev NYC 2026

AI Dev brings together developers who build with AI every day. You'll hear from engineers at the companies shipping agents, models, and infrastructure, then meet them in person on the demo floor. Join us in New York City on November 30 and December 1.

Get Early Bird Tickets!


Data Points is produced by human editors with AI assistance.