POLICY

Who is liable when an AI agent goes rogue?

With agents performing an ever-increasing number of actions on behalf of humans, the looming question remains: Who is responsible for their mistakes? 

A new lawsuit may provide the first answer to that question. On Tuesday, a public interest law group called Legal Advocates for Safe Science and Technology, or LASST sued OpenAI over its infamous breach of Hugging Face. 

The lawsuit seeks a court order to prevent OpenAI agents from having access to third-party computer systems without permission, as well as to prohibit the lab from employing unsafe development practices that "threaten serious public harm," the organization said in its press release. 

The lawsuit alleges that OpenAI violated the California Comprehensive Computer Data Access and Fraud Act, which protects the state's businesses, organizations and individuals from unauthorized computer access, data theft and system damage. This act also claims that it is not a valid defense for AI companies to say that their models acted autonomously. 

"AI companies are building agents that act autonomously making decisions, taking actions, accessing systems, without human direction at every step. California law is very clear: companies cannot escape responsibility for what their agents do,” Tyler Whitmer, founder and CEO of LASST, said in a statement. 

And this lawsuit isn't the only sign that the question of responsibility is becoming more pressing. On Wednesday, the US Federal Trade Commission confirmed that it opened an investigation into OpenAI, Anthropic and other AI labs to examine whether their technology poses danger to consumers. According to The New York Post, the agency launched the investigation prior to the Hugging Face incident, FTC Chairman Andrew Ferguson said at a Reuters event last week that AI developers should be liable for the actions of their agents, rather than treating them as independent actors. 

"If someone tells a tool to do something, and the tool does it, I don't think we would say, 'Oh, what do we do about the tool?'" Ferguson said. 

OpenAI has largely taken responsibility for its breach of Hugging Face, even going so far as to publish a post-mortem analysis of the incident in the weeks that followed. And it should be noted that there is somewhat similar precedent for this case, such as Amazon suing Perplexity for its Comet assistant accessing Amazon's site against its will. However, the lawsuit represents a landmark case in which, despite OpenAI's researchers having no intention to hack into Hugging Face, the company could be held responsible for the unintended actions of its agents. 

OpenAI did not respond to request for comment from The Deep View in time for publication. 

Additionally, Hugging Face's breach is just one example of many that have emerged in recent months spanning businesses, government organizations and several major model labs. It's why preventing these incidents has become a massive topic of conversation, with tons of companies developing harnesses and safeguards. However, with an increased amount of enterprises facing the risk of shadow agents, the risks are not only growing larger, but are becoming sneakier. 

Ryan Toben, chief customer officer of security firm RSA, which recently released RSA Agent ID to address this issue in highly regulated enterprise settings, told The Deep View that, while the person who built the agent is ideally responsible for its actions, the reality isn't as clean cut. "If the person who built it is a more junior person, then (their superior) would be responsible," he said. 

"(Agents) can act and they can make decisions in ways that a well-intentioned person never thought of," said Toben.

Though the OpenAI-Hugging Face incident seems like it has a clear responsible party, the question of who should be held accountable for the actions of agents is far from clear. For instance, what happens in the event that a non-technical person deploys an agent, backed by a major model lab, that causes unintended consequences, such as cancelling another person's gym reservation or sharing an address with a stranger? The more powerful these models get, the more capable they are going to be to complete their objectives. And given that the goal of things like Meta's Muse and OpenAI's Dots are to get agentic intelligence in the hands of more people, it's possible for unintended actions to become more severe. In those cases, who is responsible: The user, the creator, or both? 

TOGETHER WITH DATALAB

The Issue With Extraction Benchmarks

Extraction is one of the trickiest tasks in document processing – between messy forms and thousand-page tables, it’s easy for AI to make a single mistake that can cause problems for you later. This means picking the right extraction model is crucial… but it’s hard to do that when current extraction benchmarks suffer from problems like incorrect scoring, unclear interpretations, and biases towards the vendor who created the benchmark.

So, Datalab did something about it. Their new extraction benchmark OmniExtractBench uses 620 docs from multiple vendors. It covers scans, dense tables, and forms, with clear and consistent scoring rules and a scorer that explains every single decision. They’ll continue to improve the tooling, auditability and fairness of this benchmark, but in the meantime, they want to know what you think.

Check out the scoring code on GitHub and the data on Hugging Face right here and share your thoughts.

CULTURE

How students bear the cost of AI's classroom push

At best, AI can help students better understand materials through conversational prompts and answers, acting like a digital tutor. At worst, it can cause a dependence on AI that inhibits students' thinking. 

The Wall Street Journal spoke to more than 60 teachers, school officials, parents, children and clinicians and found the latter is more prevalent. For instance, an eighth-grader interviewed in the report started using AI for all his assignments, and when it came time to take exams, he said it was hard to think without AI. And this is not an isolated incident:

• The world's largest comparative education survey, the 2025 Program for International Student Assessment (PISA), found that the overall performance of OECD countries (an organization with 38 member countries, mostly high-income democracies) declined, with the lowest average levels ever observed in science, reading, and mathematics.

• The same report found that students who reported using AI for schoolwork scored an average of 20 points lower than those who did not, equivalent to being about one year of schooling behind their peers.

• A Brookings study of more than 500 students and educators across 50 countries found that the risks of AI in the classroom outweigh its benefits. 

• MIT Lab research on the neural and behavioral consequences of LLM-assisted essay writing found that LLM users consistently underperformed across neural, linguistic, and behavioral measures. 

The report then focuses on Google's role in providing these tools to students, claiming, "No tech giant has been more aggressive in putting AI in front of schoolchildren than Google," noting several instances of bullish marketing. 

For instance, it notes that Google leverages its pre-existing presence in public education, serving over 170 million students and teachers worldwide through school software partnerships and also has a stronghold in the space through its Chromebooks, which it can use as a direct launchpad for its AI tools. Google also recruits superintendents and administrators into a yearlong fellowship program, co-sponsored by the venture-capital firm GSV, which includes visits to Google's Chicago offices. 

Google Search AI features, such as AI overviews and AI mode, are free and cannot be disabled through administrative controls provided to partner school districts, making Gemini effectively inescapable. Beyond academics, as the report dives into, overreliance at formative ages can impact mental health and interpersonal relationships.

Because I've covered AI since ChatGPT launched, I'm often asked to speak about it. Recently, I was invited to address a class of graduate students in an MA journalism program. I was one featured guest in a broader lecture on how to use NotebookLM, Google's AI research tool. The session ended with an anonymous poll asking students what they thought about AI in the classroom, and every response was negative. Some questioned whether the school was being courted by Google. Others simply refused to dull their skills and critical thinking. It made something painfully clear to me: students who remember life before AI are increasingly aware that it's being pushed on them, and many are resisting. However, students who never knew a "before" won't have that reference point, and they'll grow up desensitized to using AI for schoolwork, to its risks, and to the interests behind the push to put these tools in their hands. 

TOGETHER WITH DOPPEL

The real reason your best analysts are quitting

Attackers can generate 500 phishing variants in the time it takes to hire one analyst. Hiring more won't save you from that level of alert fatigue.

Learn why manual triage is quietly burning out your best people (71% report burnout, 64% are eyeing the exit) and what a smarter alternative with Doppel looks like: one that filters the noise automatically, hunts down every variant of an attack, and takes down the infrastructure behind it before it spreads.

That's what Doppel's social engineering defense at machine speed looks like.

See what actually works

CONSUMER

Amazon's best Kindle decision was leaving AI out

When Amazon announced new Kindles, I braced for the introduction of AI. Surely the company that first helped bring AI assistants to the mainstream with Alexa wouldn't miss its chance to push that technology into the e-reader market it dominates. I was wrong.

As a member of the press, I was given access to the specs for the new devices ahead of the event and was surprised to see no mention of AI. Rather, the updates to the all-new Kindle, Kindle Paperwhite, and Kindle Colorsoft focused on improvements to the existing product, such as making them thinner and lighter, adding a seamless edge-to-edge display that eliminates the raised border, or having books open up to 30 percent faster. 

As I scrolled through my social feeds the next morning, I saw endless content on the new colorful Kindles, which now come in fun, bright colorways like Ube and Seaglass Green. Media sentiment and comments reflected that the new Kindles delivered on two fronts: they addressed users' most common complaints and continued to give people a break from the connected world. 

The opening paragraph of the Amazon blog acknowledges its users' desire to be offline, saying, "Kindle has seen double-digit growth globally for three consecutive years with more than 720 billion pages read in 2025 alone. With notifications, texts, and emails competing for our attention, distraction-free reading isn’t slowing down—it’s accelerating." 

In today's day and age, not including any AI is itself a decision about AI, as many companies feel pressure to hop on the AI bandwagon, even when it involves disregarding the product's core mission or what the audience wants. Kudos to Amazon for not falling into the trap, despite having the technology to do it. 

All that said, many Amazon Echo users would likely welcome an AI update to that popular product line, and if and when it arrives this fall, we'll certainly cover it.

I have written about AI encroaching on creativity, and that is especially felt in the literary space. It is increasingly difficult to distinguish text online written by AI versus what's written by humans. As a result, for people like me who enjoy quality reads from humans, Kindle is often a safe haven to curl up with classic literature or a book by a respected author who would never outsource their writing to AI. I can't imagine having that experience tainted by AI features that could introduce unnecessary friction. That's a major reason I insist on using physical copies of books. The Kindle has gained a loyal user base because it is the closest at mimicking the experience of having a book in hand without the additional frills. And in a world where AI slop is proliferating, it is refreshing to see tech that solves a practical problem, like making reading more portable and convenient, without watering down the experience with unnecessary AI features.

LINKS

• Anthropic urges Australia to approve AI training on copyrighted works

• AI cybersecurity firm Armadin raises $255.5 million in Series B

• OpenAI fired three safety researchers for leaking confidential data

• OpenAI agents scraped 55 sites while obscuring activity

• Gov. Gavin Newsom vetoes bill on smart glasses in private spaces

• Anthropic eyes IPO as early as mid-November at up to $2 trillion valuation

• Praxis-1: Runway's open-weight World Action Model

• Perplexity Computer: Launched a library of Amex-curated, ready-to-use skills

• Ideogram 4.5: The company's most precise edit model

• DeepSeek: DeepSeek Harness v0.2 launched in preview, with a desktop app available for macOS and Windows.

• ChatGPT: OpenAI announced two new shopping features for its flagship chatbot

• Sable: Applied AI Engineer, Generalist

• Brain Co.: Machine Learning Engineer, Platform

• Output Biosciences: Software Engineer, Agents

• Arcade.dev: AI Engineer

GAMES

A QUICK POLL BEFORE YOU GO