March 10, 2026

Codex security agent, Claude fights back & more

Codex security agent, Claude fights back & more

Today’s Overview

Enterprises are rapidly adopting AI tools that blend productivity, security, and multimodal capabilities, while major vendors deepen their ecosystems and confront regulatory scrutiny. OpenAI, Anthropic, Google, and others are launching new models, integrations, and security agents that accelerate AI‑driven workflows and raise the stakes of compliance.

  • OpenAI broadened its GPT-5.4 suite with the Codex Security autonomous agent and a ChatGPT for Excel add-in, targeting enterprise code safety and spreadsheet automation.
  • OpenAI’s acquisition of Promptfoo brings an open-source security, evaluation and compliance platform into its model pipeline, strengthening AI product testing.
  • Anthropic added a memory-import feature to Claude, allowing users to transfer personalization data from competitors in under a minute and reducing switching barriers.
  • Google rolled out new Gemini AI capabilities across Docs, Sheets, Slides and Drive and open-sourced the Always On Memory Agent, enhancing continuous context for enterprise users.
  • Tencent AI Lab unveiled Penguin-VL, a compact vision-language model with a Penguin-Encoder that improves multimodal reasoning efficiency for business applications.
  • Anthropic filed lawsuits against the Pentagon blacklist and a White House directive, underscoring growing regulatory challenges for AI firms.

Top Stories

OpenAI broadens GPT-5.4 suite with Codex Security agent and ChatGPT for Excel add-in

OpenAI expanded its GPT-5.4 ecosystem with the launch of Codex Security, an autonomous agent that scans entire code repositories, builds project-specific threat models, and cuts false positives by half. In its first 30-day beta, the agent examined 1.2 million commits, identified 792 critical vulnerabilities and helped secure 14 CVEs for projects such as OpenSSH, GnuTLS and Chromium. OpenAI also released a beta ChatGPT for Excel add-in that lets users create, modify and analyze spreadsheets using natural language while preserving formulas and cell structures, and it integrates data from providers such as FactSet and S&P Global. Additionally, the company announced a program for open-source maintainers that provides six months of free ChatGPT Pro access, API credits and the Codex agent, complementing its one-million-dollar Codex Open Source Fund.

Read Full Article

Promptfoo to be acquired by OpenAI

Promptfoo announced that it will be acquired by OpenAI while remaining an open-source project. Founded in 2024, the platform helps developers systematically test AI applications for security, evaluation and compliance. OpenAI plans to integrate Promptfoo’s testing technology into its models and infrastructure to enable teams to detect vulnerabilities early and ship more secure AI products.

Read Full Article

Anthropic sues U.S. government over Pentagon blacklist and White House directive

Anthropic filed lawsuits in two federal courts challenging a Pentagon blacklist label applied to its technology and a White House order that would prohibit federal agencies from using the Claude model. The company says the actions retaliate against its public advocacy for AI safety limits on weapons and surveillance. An amicus brief signed by more than 30 former OpenAI and Google engineers supports Anthropic, warning that the blacklist could undermine U.S. AI leadership. The cases could determine whether the government may penalize a domestic AI firm for safety-related speech.

Read Full Article

Research & Analysis

Google researchers enable LLMs to perform Bayesian reasoning

Google researchers developed a method to train large language models to execute Bayesian inference, allowing them to produce optimal probabilistic predictions. The approach significantly improves LLM performance on recommendation tasks, where accurate probability estimates are critical. Experiments show that the Bayesian-trained models generalize better to a range of downstream tasks beyond recommendation. These findings suggest that incorporating formal probabilistic reasoning can enhance the reliability and versatility of LLMs.

Read Source

Trending Tools

  • Google PM releases open-source Always On Memory Agent

    Google’s product manager open-sourced the Always On Memory Agent, which continuously captures and consolidates data for later retrieval without relying on traditional vector databases. The MIT-licensed project helps enterprise AI teams maintain context and user preferences across extended interactions.

  • Andrew Ng launches Context Hub for AI agents

    Andrew Ng introduced Context Hub, an open-source utility that supplies AI coding agents with up-to-date documentation and reference material. By integrating the tool into development workflows, agents can retrieve accurate information, reducing hallucinations and ensuring code suggestions reflect the latest APIs.

  • Tencent AI Lab unveils Penguin-VL vision-language model

    Tencent AI Lab released Penguin-VL, a compact family of vision-language models featuring the Penguin-Encoder that aligns visual features with language using a text-only LLM as a foundation. The design improves multimodal reasoning efficiency while keeping model size small.

Quick Hits

Join the AI Recap Newsletter

Get the latest AI news, research insights, and practical implementation guides delivered to your inbox daily.

By subscribing, you agree to our Terms of Service and Privacy Policy.