AI Daily: Sora App Shutdown, Claude Auto Mode, and LiteLLM Security Breach

The Evolution of Sora’s Service and Agent Tools

Observing recent trends in the tech world reveals several unexpected turns. Products that were expected to follow a set path have suddenly shifted gears. From the exit of video generation apps to developer tools gaining more autonomy, these events map out the trajectory of a tech industry moving toward maturity and systematization. What does this really mean? Let’s dive into these major developments and see what has happened over the past few weeks.

Sora App Officially Bids Farewell: OpenAI Exits Video Generation Market

As many know, it was only at the end of September 2025 that the video generation tool, which amazed countless creators, launched its standalone app. However, the official Sora team recently announced that it will be shutting down this application service. In a statement, the team thanked all users who used the tool to create and build a community, acknowledging that this news might be disappointing to many. The team promised to release a timeline for the app and API shutdown soon, along with details on how creators can save their work.

This is undoubtedly a bombshell. According to reports from The Hollywood Reporter, OpenAI has decided to completely exit the video generation business. This decision directly impacted the plans of entertainment giant Disney. Disney had pledged an investment of up to one billion dollars in OpenAI at the end of last year, with plans to license some famous characters for the platform. That massive deal has now collapsed.

A Disney spokesperson provided a formal response, stating that as the nascent AI field flourishes, Disney respects OpenAI’s decision to exit the video generation business and shift its focus elsewhere. The spokesperson emphasized that Disney is grateful for the constructive collaboration between both teams and the experience gained, and will continue to participate in various platforms to find responsible ways to embrace new technology while ensuring respect for intellectual property and creator rights.

This move indicates a clear industry trend. As base model developers decide to focus their energy back on core logic and text models, the landscape of the video generation market is bound to shift. This also confirms that the business models for the application layer and base models are undergoing a brutal market restructuring.

According to the latest report from The Wall Street Journal, OpenAI CEO Sam Altman has explicitly informed employees that the company will phase out all products using its video models. This includes not only shutting down the consumer Sora standalone app but also terminating the Sora API version for developers, and even stating that “video features will not be supported in ChatGPT.”

Claude Introduces Auto Mode: Giving Decisions to the System

While video tools are reshuffling, tools in the software development field are seeing interesting upgrades. The Anthropic team recently launched a new Auto Mode for its development tools, a feature that will surely catch the eye of many engineers.

In the past, when using such coding assistants, developers often faced a dilemma: either constantly manually approve every file write and terminal command, or bypass all permission checks entirely. Bypassing checks is convenient but comes with unpredictable risks. Now, this new mode offers a very smart middle ground.

The principle is quite intuitive. Before each tool call, an internal classifier evaluates whether the action is destructive. If the classifier deems the action safe, the system proceeds automatically. If potential risks are found, such as mass file deletion or unauthorized data transfer, the system will block the action and guide the program to try other safe solutions.

Of course, risk management is never perfectly foolproof. The team also reminded users that while this mechanism reduces risk, it cannot completely eliminate all hazards. It is strongly recommended that users execute these automated tasks in an isolated sandbox environment. Currently, this feature has been launched as a research preview in the Team plan, and Enterprise and API users will receive updates in the coming days. This step marks the transition of the system from a passive execution tool to an intelligent agent with autonomous judgment.

Long-Running Applications: The Endurance of Agent Systems

Speaking of autonomous decision-making, it’s worth discussing how to keep these smart systems stable during “continuous overtime.” Anthropic’s engineering team recently shared an article on architectural design for long-running applications, exploring challenges that are very close to reality.

Honestly, getting a system to run continuously for hours and produce valuable code is extremely difficult. When processing vast amounts of information, models often experience “context anxiety”—as memory fills up, the system rushes to finish the job, leading to a significant drop in quality. To solve this bottleneck, the engineering team took inspiration from Generative Adversarial Networks (GANs) and designed a multi-agent architecture consisting of a Planner, a Generator, and an Evaluator.

The Planner is responsible for breaking down large goals into small tasks, the Generator focuses on writing code, and the Evaluator plays the role of quality assurance. The Evaluator even operates a browser like a human to test the interface for issues. This approach of subdividing work and establishing a feedback loop has successfully enabled the system to autonomously write complete web applications, including both front-end and back-end components.

For a real-world analogy, think of running a restaurant. The Planner is the head chef who creates the menu, the Generator is the cook who preps and cooks the food, and the Evaluator is the picky taster. Only when the taster nods can the dish be served. This architecture provides a valuable framework for future automated operations or long-running workflows.

LiteLLM Suffers Supply Chain Attack: Challenges for the Open Source Ecosystem

With great power comes great risk. While everyone was cheering for new features, a serious cybersecurity incident also broke out. The LiteLLM project on PyPI was hit by a supply chain attack, immediately putting the industry on high alert.

LiteLLM is a popular tool used to unify calls to multiple large language model APIs, and many applications rely on it. However, in version 1.82.8 released on March 24, malicious code was injected. Once the infected version is installed, the malware runs silently every time the Python environment is started.

The malware’s behavior is quite aggressive. It not only harvests SSH private keys, cloud service credentials, and database passwords from the host but also attempts to create privileged backdoors in Kubernetes clusters. Interestingly, the team that discovered this vulnerability did so because the malware had a small bug that caused the system to repeatedly start child threads, eventually crashing the memory and getting noticed.

This incident is a lesson for everyone. As related applications become more popular, the high dependency of development environments has become a target for hackers. This suggests that package verification and least-privileged access control will become incredibly important in future development processes. It’s a repeat of the security issues from the early days of the internet, just on a new battlefield.

OpenAI Foundation Commits $1 Billion to Tech and Social Challenges

After looking at the tug-of-war between technology and security, let’s focus on the progress of tech influence extending to the social level. Technological progress ultimately returns to solving real human problems. The OpenAI Foundation recently released its latest updates, showcasing specific plans to promote social good.

The statement noted that the foundation expects to commit at least one billion dollars in resources over the next year. These funds will be primarily used in several key areas. First is life sciences and disease treatment, particularly for complex and heartbreaking diseases like Alzheimer’s. The foundation plans to work with top research institutions, using powerful data analysis capabilities to find new biomarkers, hoping to accelerate the development of treatment options.

Second is employment and economic impact. Machine learning technology will inevitably change the landscape of the labor market. The foundation commits to working with civil society, unions, and economists to develop practical solutions to help workers adapt to change. Finally, system safety and social resilience, including protecting children from inappropriate content and preventing biosecurity threats.

As software systems grow in capability at an alarming rate, ensuring that these powers can safely and fairly benefit the public is just as important as writing an elegant line of code. This is a serious issue that the entire industry must face together.

FAQ

To help clarify the details of these industry changes, here are some of the most common questions.

Q: Why is the Sora app closing? A: OpenAI has decided to exit the video generation business entirely and shift company resources and priorities to other core areas. As a result, the standalone app launched last September will cease operations. The official team will soon announce a detailed timeline for work backups and the API shutdown.

Q: What is the impact of Disney canceling its investment in OpenAI? A: Disney originally planned to invest one billion dollars and license character IPs. Now that the deal is canceled, Disney will temporarily turn its attention to other technical platforms that better suit its business interests. This has also significantly changed the competitive landscape of the video generation market.

Q: How should developers respond to the LiteLLM security vulnerability? A: If you installed or upgraded LiteLLM after March 24, 2026, be sure to check if you have the infected version 1.82.8 or 1.82.7. If installed, you must immediately remove the package and clear the cache. It is also strongly recommended to change all keys and password credentials on that host, as this data may have already been leaked.

Q: Is Claude’s Auto Mode really safe enough? A: Auto Mode has a built-in protective classifier that automatically blocks high-risk commands like mass deletion or data leakage. However, the official team admits this cannot prevent all dangers 100%. The best practice is always to run such automated tasks in an independent, isolated environment for an extra layer of security.