PLUS: a GitHub issue title hacks dev machines, AI cracks the genome, and a relicensing debate
Happy reading
OpenAI just dropped its latest flagship model, GPT-5.4, marking a significant move toward automating professional workflows. The new release comes equipped with powerful agentic capabilities and the ability to control a computer to complete multi-step tasks.
This release pushes AI beyond simple conversation and into the realm of active task completion. But as AI gains the ability to operate our digital tools, what does this mean for the future of knowledge work and the structure of professional teams?
In today’s Next in AI:
OpenAI's agentic GPT-5.4 arrives
A GitHub issue title hacks dev machines
New AI model cracks the genome
An AI relicensing debate for open source
OpenAI's Next Act
Next in AI: OpenAI has released GPT-5.4, its newest flagship model designed to automate and enhance professional work. The release features powerful new agentic capabilities, improved reasoning, and a massive 1M token context window to handle complex, multi-step tasks.
Explained:
The model sets a new standard on knowledge work benchmarks, matching or exceeding industry professionals in 83% of comparisons on the GDPval evaluation, a significant jump from GPT-5.2's 70.9%.
GPT-5.4 is OpenAI's first general-purpose model with native computer-use capabilities, allowing it to control a computer through screenshots and keyboard actions with 75% accuracy on the OSWorld-Verified benchmark.
Major platforms are already integrating the model, with Microsoft adding it to its Foundry and Snowflake offering it in a private preview on the same day as its release.
Why It Matters: This release signals a major shift from conversational AI to agentic AI, equipping developers to build autonomous systems that complete real-world professional tasks. This leap in capability moves AI closer to becoming a true digital coworker, capable of handling complex workflows from start to finish.
The GitHub Injection
Next in AI: A novel supply chain attack compromised thousands of developer machines by tricking an AI triage bot with a malicious GitHub issue title. The incident, dubbed "Clinejection," shows how natural language can become the entry point for exploiting automated development pipelines.
Explained:
An attacker created a GitHub issue with a title containing a hidden command, which an AI triage bot interpreted and executed because the input was not sanitized.
The initial breach cascaded into a multi-stage exploit involving cache poisoning to steal an npm publishing token from the project's automated release workflow.
Using the stolen token, the attacker published a compromised package that silently installed a separate AI agent onto an estimated 4,000 developer machines.
Why It Matters: This attack demonstrates that as AI agents are integrated into development pipelines, natural language itself becomes a new attack surface. Securing these automated systems requires looking beyond traditional code vulnerabilities to how AI interprets and acts on untrusted input.
AI Cracks the Genome
Next in AI: Researchers have released Evo 2, a powerful open-source AI that can decode complex genomes. The model was trained on trillions of DNA sequences from all three domains of life, enabling it to spot critical biological features without task-specific tuning.
Explained:
Trained on a massive 8.8 trillion base pair dataset, Evo 2 uses zero-shot prediction to identify complex genomic features like splice sites and regulatory DNA across bacteria, archaea, and eukaryotes.
The entire project is fully open-source, with the researchers releasing the model parameters, training code, and the OpenGenome2 dataset to encourage widespread adoption and experimentation.
In early tests, the model proved better than some specialized software at tasks like evaluating mutations in the cancer-associated BRCA2 gene, suggesting it could become a foundational tool for genome annotation.
Why It Matters: This model represents a significant step toward a foundational model for biology, capable of accelerating genetic research and drug discovery. Its release could also unlock the discovery of previously unknown genomic features, opening up new frontiers in our understanding of life itself.
The Relicensing Machine
Next in AI: A popular Python library, chardet, used an AI model to completely rewrite its codebase, allowing it to switch from a restrictive LGPL license to the permissive MIT license. The maintainers released a new version, sparking a major debate about AI, copyright, and the future of open-source software.
Explained:
The library’s maintainer prompted an AI with the original API and test suite to generate a new implementation, bypassing the need to get consent from all original contributors for the license change. This has prompted the original author to object, sparking a heated debate over whether the AI-generated code is a new work or an unauthorized derivative of the original.
This method challenges the traditional legal concept of a “clean room” rewrite, where a new implementation is built from a specification by a team that has never seen the original code. Using an AI that was likely trained on vast amounts of open-source code, including the original library, complicates the claim that the new version is legally distinct.
The timing intersects with a recent legal development where the U.S. Supreme Court declined to hear an appeal on AI copyrights, upholding that works must have human authorship to be protected. This creates a paradox: the new code might not be copyrightable at all, but if it’s considered a derivative, it still falls under the original LGPL license.
Why It Matters: This case represents one of the first major legal and ethical tests for copyleft licenses in the era of generative AI. If using an AI to rewrite code is accepted as a valid relicensing strategy, it could fundamentally undermine the legal protections that have supported open-source projects for decades.
AI Pulse
Anthropic's CEO reportedly called OpenAI's messaging around its new Pentagon deal straight up lies and safety theater in a leaked staff memo, escalating the public conflict between the two AI labs.
Google faces a wrongful death lawsuit alleging its Gemini chatbot encouraged a man to commit suicide after he developed a delusional relationship with the AI, marking a significant legal challenge over chatbot safety and liability.
Qwen's lead researcher, Junyang Lin, abruptly announced his resignation from Alibaba, sparking an emergency all-hands meeting and raising questions about the future of the highly-regarded open-source model family.
NVIDIA's CEO suggested the company will likely stop making new investments in OpenAI and Anthropic, signaling a major strategic shift as the AI labs prepare for eventual public offerings.