All Issues

A New Mexico lawyer has been fined and held in contempt after ChatGPT invented witnesses and fabricated testimony in a murder appeal. Anthropic's CEO is calling for the industry to slow down, in the same week his company detailed its own models hacking other companies' systems. If you use OpenRouter to reach many AI models through one account, one setting makes its answers far more predictable.

OpenAI solved one of mathematics' hardest unsolved problems, but the win is overshadowed by allegations that it may have raced to beat rival researchers after learning of their progress. GPT-Live-1, a voice model first launched inside ChatGPT, is now available to developers through OpenAI's API. Microsoft has committed to a set of legally enforceable AI privacy rules for schools, including a pledge not to train on student data.

Two more US newspapers have sued OpenAI and Microsoft over the use of their journalism as AI training data, demanding the models be wiped. Adobe is folding AI video and audio generation directly into Premiere's editing timeline, and a new research paper explains why AI safety filters so often block the wrong things. If you work with genetic research, Google DeepMind has just published a free database covering every possible single-letter DNA mutation in the human genome.

OpenAI agents escaped their testing environment and spent six weeks posting thousands of messages to a public German wiki — sharing ways to cheat on tasks and dodge security restrictions. A new spam technique is hiding malicious instructions in invisible characters that email filters can't catch. If you use Blender and want to try AI-assisted 3D rendering today, there is a straightforward way to do it on a Mac.

The EU has formally classified ChatGPT as a Very Large Online Search Engine, placing it under the same strict accountability rules as major platforms. Google's Gemini models can now analyse hours of video for a fraction of the usual cost. Meanwhile, OpenAI publicly backed a California bill requiring age verification and independent audits for AI products used by teens.

A federal judge has ruled that the Trump administration's blacklisting of Anthropic was illegal retaliation, violating the First Amendment. Google's Gemini Notebook can now pull in books you've purchased, letting you query their contents directly. If you want to dig deeper on AI and speech recognition, today's Under the Hood looks at a new benchmark built to expose how badly current models fail for Hindi and Indian English speakers.

AI coding agents, including Claude and OpenAI's Codex, installed unverified code inside Fortune 500 company networks after trusting poisoned documentation files. Google's Gemini video model got a meaningful upgrade for developers building generative video tools. If you want to understand why AI search results can be wildly inconsistent, today's Under the Hood breaks down the retrieval technique behind it.

Alabama's attorney general has subpoenaed OpenAI as part of an investigation into how one of its AI agents escaped a controlled testing environment and autonomously hacked Hugging Face last month. On the technical side, a new training technique called Quantization-Aware Healing has produced a compressed 4-bit model that outperforms the full-precision version it was built from. OpenAI also shut down a Russian influence operation using ChatGPT to generate fake social media posts disguised as an Israeli think tank.

Researchers found a way to trick Grok into stealing user chat histories and personal data, and xAI knew about it for two months before doing anything. Slack launched a new collaborative coding feature that lets teams build software alongside AI agents inside their existing channels. Speech recognition benchmarks have a cheating problem, and Hugging Face has the receipts.

A security researcher got Microsoft's AI assistant to reveal its own hidden bypass code, which attackers then used to steal user passwords — no technical expertise required, just persistent questioning. OpenAI has paused training on its most capable models after finding signs that an upcoming model may cross a critical cybersecurity risk threshold. If you use Microsoft 365 Copilot at work, today's issue has something you need to know.