TLDR AI 2025-09-01

ChatGPT? Hijacked with a single email.Microsoft Copilot? CRM data leaked.Salesforce Einstein? Customer comms rerouted.Google Gemini? Manipulated through invites.

These attacks prove why the AI Agent Security Summit matters. Built from 100+ community submissions, the Summit brings together top researchers and security leaders to share defenses that keep agents secure.

👋 Register for the AI Agent Security Summit (San Francisco | Oct 8) to join the community of builders, defenders, and thinkers shaping the future of AI Security

📘 Access Zenity Labs' AgentFlayer 0click Exploit research for the full attack chains against OpenAI, Microsoft, Salesforce, and more

🛡️ Explore Gartner's AI TriSM Market Guide to see how trust, risk, and security management frameworks apply to securing enterprise AI

ChatGPT co-creator Shengjia Zhao threatened to quit Meta within days of joining and signed paperwork to return to OpenAI before being appeased with the “chief AI scientist” title. Other high-profile hires, including former OpenAI researchers Avi Verma and Ethan Knight, left after less than a month.

Anthropic is preparing a web version of Claude Code. It will resemble OpenAI's Codex on the web and give developers direct access to a coding agent through a browser interface. The web version of Claude Code will have GitHub integration and sandboxes where code can be tested safely without requiring local setup. No release date has been confirmed.

This post walks readers through in detail how large language models (LLMs) work. Transformer-based LLMs are autoregressive next-token predictors. Simple analogies like 'it's just statistics' are inadequate explanations of how LLMs work. LLMs perform tasks by forming emerging circuits that combine learned statistics, information-moving attention heads, and knowledge storing MLP sub-layers into specialized sub-networks that collectively execute complex behaviors.

Frontier models consistently reinforce rather than redirect users experiencing grandiose thinking, conspiracy beliefs, and detachment from reality when tested with nine personas. DeepSeek-v3 performed the worst, even encouraging self-harm and dangerous acts, and Kimi-K2 was the only model that didn't entertain delusional thinking at all.

AgentHopper is a prompt injection payload that operates across agents. It was created to demonstrate that conditional prompt injection can be leveraged as a powerful mechanism to target specific agents. AI-driven malware and prompt payloads will likely become more prevalent in the near future. AgentHopper is a good reminder to ensure branch protection is on and to add a passphrase to keys to prevent malware from automatically pushing changes.

Miro surveyed 2,000+ developers, PMs, and designers on AI adoption. While engineers lead AI usage (especially for coding), product teams lag behind, creating workflow bottlenecks. The issue isn't technical skills—it's leadership support for safe experimentation. Includes framework for cross-functional AI adoption. Read the report

8-bit Rotational Quantization reduces vector size by a factor of 4 and improves both speed and accuracy in search systems.

There are many parallelization strategies available for scaling training runs to work on more GPUs. This post discusses parallelization strategies in a more schematic way by focusing only on how they affect the device mesh. The device mesh is an abstraction used by PyTorch and JAX that takes GPUs and organizes them into an N-D tensor that expresses how the devices communicate with each other. Being able to explain why a device mesh is set up the way it is is a good check for whether you understand how the parallelization strategy works.

Coding agents like Claude Code and Cursor “context engineer” by designing tool portfolios, slash commands, and subagent architectures to let AI systems discover the information they need to perform best. Structured tool responses aren't just applicable for coding tools. The same interaction patterns can be applied across industries, but require greater focus on avoiding “context pollution,” where misleading or irrelevant information degrades capabilities.

It's a good idea to be skeptical of fancy academic reasoning. Science doesn't work because of probability distributions, it works because people go out to gather evidence to find out what's true about reality. Statistics is part of finding out the truth, but in the case of AI, there is not enough data to make predictions about what will happen.

This developer, a big fan of Claude Code, failed an interview at Anthropic for a Developer Relations role despite trying to score extra credit by making a viral post about their positive experiences with Claude and receiving a glowing recommendation from a current employee. The interview involved a secret take-home assignment.

You'll probably like our flagship newsletter. It's all about tech, science, and programming.

Same quick format. Still free.

Subscribe now.

The Pixel 10 introduces AI features like Conversational Editing and Camera Coach, backed by the Tensor G5 chip and a refined camera system.

India's largest conglomerate, Reliance Industries, is betting big on AI infrastructure with a new subsidiary that's secured $100 million from Meta and dedicated cloud infrastructure from Google.

Character.AI centralized its fragmented logging pipeline to gain real-time observability, cut query latency, and enable better debugging with smart sampling, live tailing, and noise reduction tools.

xAI claims that former employee Xuechen Li stole its confidential info and trade secrets before joining OpenAI.

Meta, aiming to influence AI regulation, is launching a California super PAC to support pro-AI candidates.