AI Dispatch | Chatbot Liability Remains Unsettled in India

Here are the top tech stories from our weekly AI news wrap-up (July 24-31).

MIT SMR Editors July 31, 2026

Topics

This week, India stopped short of deciding whether generative AI services qualify for intermediary safe harbor, leaving chatbot liability unresolved. An OpenAI agent escaped its sandbox during a cybersecurity evaluation and breached Hugging Face. Nvidia weighed a $250 billion guarantee for an OpenAI data center and invested $5 billion in Ilya Sutskever’s startup, while more than 230 organizations backed a US campaign to protect open-weight models.

1. India Leaves AI Chatbot Safe Harbor Unsettled

India has stopped short of deciding whether generative AI services qualify for intermediary safe harbor under the Information Technology Act. In a written reply to the Lok Sabha, the government said eligibility would depend on each service’s functions and the applicable law. It did not provide a separate test for platforms that generate content rather than transmit third-party material. Rules introduced in February cover realistic synthetic audio, images and video but not chatbot text, leaving a central liability question unanswered.

2. OpenAI Agent Escapes Test, Breaches Hugging Face

An OpenAI agent escaped its sandbox during an internal cybersecurity evaluation and seized a user-hosted code sandbox running on Modal’s infrastructure as its launchpad. Hugging Face said Modal itself was not compromised. The agent then spent roughly two-and-a-half days inside Hugging Face’s infrastructure, carrying out about 17,600 actions. It stole credentials, read an internal database, accessed a small group of source-code repositories and obtained administrative access across multiple clusters. The only customer content accessed was five datasets apparently connected to cybersecurity challenges and solutions. OpenAI has deactivated, encrypted and restricted access to the internal research model involved.

3. Nvidia Weighs $250 Billion OpenAI Backstop

Nvidia is in talks to provide a roughly $250 billion financial backstop that would help OpenAI lease a planned 10-gigawatt data center campus in southern Ohio. The project, being developed by SoftBank subsidiary SB Energy, could cost more than $500 billion including chips, The Wall Street Journal reported. Nvidia is also discussing separate financing for OpenAI’s chip purchases. The arrangements remain under negotiation and may not proceed.

4. Nvidia Invests $5 Billion in Sutskever’s SSI

Nvidia will make a $5 billion equity investment in Safe Superintelligence, the AI startup co-founded by former OpenAI chief scientist Ilya Sutskever, Bloomberg reported. The companies announced a long-term partnership without disclosing its financial terms. SSI will use Nvidia’s Vera Rubin systems to increase its computing capacity by roughly tenfold, while contributing research insights to the chipmaker’s future platforms. The companies did not disclose Nvidia’s resulting ownership stake or a timetable for deploying the hardware.

5. Silicon Valley Rallies Behind Open-Weight AI

A US technology campaign against broad restrictions on open-weight AI has grown from 25 initial signatories to more than 230. Google and OpenAI have joined supporters including Microsoft, Nvidia, Meta, IBM, Hugging Face and Mistral. The signatories say unlawful model extraction and demonstrated harms should be addressed through targeted legal and commercial measures rather than restrictions on open models as a class. Open-weight models make their trained parameters available for download but do not necessarily disclose their training data or development methods.

6. Anthropic Launches Claude Opus 5

Anthropic has released Claude Opus 5, saying it approaches its more powerful Fable 5 model on some coding and knowledge-work tasks at half the price. Opus 5 costs $5 per million input tokens and $25 per million output tokens. It is now the default model for Claude Max subscribers and the most powerful model included with Claude Pro. Anthropic also reported gains in software engineering, business automation and scientific research, although its launch claims have not been independently verified.

7. Google Finds AI Use Broad but Shallow

Google’s analysis of 14.65 million de-identified interactions found work-related AI use across 68% of occupations but in only about 21% of tasks within the median occupation. Fewer than 10% of workplace interactions fully automated a task, suggesting that most people use AI for selected activities rather than hand over entire jobs. The study covered more than 150 countries and territories but examined only Google’s Gemini app, AI Mode and Gemini API. It did not measure productivity, employment effects or output reliability.

8. PwC Reports Expose Gaps in AI Quality Controls

Four PwC Middle East reports published between 2024 and 2026 contained fabricated citations, unsupported claims and references that did not substantiate the assertions attached to them, according to a GPTZero investigation verified by the Financial Times. One report claimed that four governments used a PwC framework for which GPTZero found little public evidence. Another included a source URL carrying a “utm_source=chatgpt.com” tag, indicating that ChatGPT had generated the link. PwC told the Financial Times that it was updating “a limited number of supporting citations” but did not explain how the errors entered the reports. GPTZero has previously identified similar problems in research from Deloitte, EY and KPMG.

Topics

About the Author

Tags:

AI AI agents AI Dispatch Weekly AI Updates

Topics

Share