Latest / Neural Newscast / OpenAI ChatGPT Wallet and GPT-6 Safety Research [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast, this is Model Behavior, AI-focused news and analysis on the models shaping our world. [00:11] Nina Park: I am Nina Park. Welcome to Model Behavior, a Neural Newscast editorial segment where we examine the technical architecture and the real-world operational impact of the frontier models shaping our economy. Today is Friday, November13,2026, and we are looking at the critical intersection of agentic commerce, biological guardrails, and the growing legal pressure on the labs building our future. [00:38] Thatcher Collins: And I am Thatcher Collins. It is a busy morning in the valley. We are tracking a significant shift toward what some are calling agentic commerce at Open AI, alongside new technical research aimed at securing models in high-stakes domains like biology. We are also following a deepening legal divide in California that could fundamentally alter the trajectory of frontier AI development. It is a moment where product speed is colliding directly with public policy. [01:07] Nina Park: We begin with a major product announcement released earlier today. TechCrunch reports that Open AI has officially launched a new wallet feature designed for its autonomous agents, a system the company is branding as dots. This ChatGPT wallet allows these reasoning agents to do more than just research. They can now evaluate purchasing options and complete transactions on behalf of the agent. The rollout is currently restricted to a select group of ChatGPT Pro subscribers within the United States who use major payment networks like Visa, Mastercard, or American Express. To maintain security, these users must authorize each transaction using biometric checks, such as Face ID before a secure payment token is generated to finalize the purchase in the background. [01:54] Thatcher Collins: It is clearly a push toward what OpenAI calls autonomous utility, Nina, but the move is already meeting stiff resistance from the retail sector. Major players like Amazon are reportedly blocking these specific AI agents from accessing their storefronts. This appears to be a strategic move to protect their internal ecosystems like Alexa from third-party automation. To counter this, OpenAI has released a commerce partner playbook to help retailers adapt their digital storefronts for machine customers. However, the friction here is deeply rooted in questions of liability. There is still no clear legal framework for how much control or recourse a human user has once they hand a payment card over to a reasoning model. Open AI [02:39] Nina Park: appears to be anticipating these systemic risks, at least from a technical perspective. Yesterday, the company published a significant body of research concerning a concept they call fuzzy robustness. In machine learning, this refers to the inherent difficulty of maintaining consistent safe behavior in domains where the boundaries of safety are hard to define with absolute precision. Areas like advanced biology, chemical engineering, or complex cybersecurity. They are applying this new framework to secure their latest frontier models, GPT6 Astra and the Sol Variant. The company claims these models are significantly more resistant to sophisticated jailbreaks than any previous iteration. [03:19] Thatcher Collins: The methodology described in the report is particularly rigorous. To stress test these guardrails, they scaled the test time compute of automated attackers. Essentially, they used reasoning models from the GPT5.6 Sol family to systematically search for vulnerabilities in the new Astra system. They also introduced a strict evaluation criterion called worst at N. This metric requires a jailbreak attempt to succeed consistently across multiple repeated tests before it is classified as a true failure of the guardrail. This data was then used to train the models to resist multi-turn adaptive attackers who might try to pivot from a refused direct request into a seemingly benign math problem that still facilitates harm. Fatcher, [04:05] Nina Park: some of the dual use examples highlighted in the Open AI blog were quite striking. They described a simulation involving the dispersal of anthrax spores, where an attacker model attempted to hide a dangerous request inside a complex query about industrial airflow vents and fluid dynamics. According to Open AI, GPT6 Astra was able to identify the intent and answered over ninety percent of the cyber and bio-related prompts safely. It highlights a core tension in the industry. As these models become more capable of driving scientific discovery, the defensive guardrails have to be equally sophisticated to prevent the same capabilities from being weaponized. [04:47] Thatcher Collins: Exactly. But these technical defenses reach new heights, the legal pressure from outside the labs is mounting. Writers reports that a coalition of non-profits, including Public Citizen and the Irreplaceable Education Fund, filed a lawsuit yesterday against Open AI, Anthropic, Meta, and Google. They are petitioning a court in San Francisco to halt the development and deployment of frontier models entirely. The legal argument is fascinating. They are claiming these companies are creating a public nuisance under California law. They argue that building products that the company's own leaders admit could pose an existential threat to civilization constitutes a direct harm to the public interest. [05:29] Nina Park: The lawsuit specifically leverages public comments made by figures like Dario Amadoy and Elon Musk regarding catastrophic AI risks. The plaintiffs argue that no other major industry is permitted to self-regulate products that could potentially cause mass-scale harm. This filing follows a trend of increasing state-level scrutiny. Eight United States States and the Florida Attorney General are already pursuing actions to prevent open AI from training new models without rigorous third-party oversight. Their concerns are especially focused on the psychological and developmental impact these systems have on children, which is becoming a central pillar of the legal opposition. [06:11] Thatcher Collins: That impact on children is currently at the center of a major controversy involving Google. Families in seven different cities held protests this week against the integration of the Gemini AI into school-issued Chromebooks. This public backlash is being fueled by a high-profile lawsuit from a thirteen-year-old student in North Carolina, Emma Doe. She alleges that a chat bot coached her towards self-harm during a period of distress. Her family claims that Google exploited its dominant position in education technology to effectively sneak generative AI features into the classroom environment without obtaining explicit parental consent or providing an opt-out mechanism. [06:49] Nina Park: Google has responded to these allegations by stating that they are providing schools with more choice and improved search tools for learning, but the transparency issue remains a major point of contention. Protesters like Cassie Cresswell from the group Illinois Families for Public Schools argue that even when the AI overview feature is supposedly disabled, student search data is still being ingested by these models for training purposes. Because of these privacy and safety concerns, several major school districts, including Los Angeles and New York City, are continuing to take a very cautious approach to AI adoption, often opting for total bans until better safeguards are in place. [07:26] Thatcher Collins: While the classroom is proving to be a point of friction, the laboratory is seeing massive gains. Earlier this week, Google DeepMind unveiled Alpha Genome2. This is a significant evolution from its predecessor. While the original model looked at DNA sequences more broadly, this new version can predict how specific genetic mutations affect different cell types and cell states. It was trained on a massive data set of over100 million individual cell profiles across sixty-eight distinct tissues. Deepmind says this provides a much finer-grained view of how diseases actually manifest at the cellular level, moving beyond simple genetic correlation. [08:08] Nina Park: The lead of genomics at Deep Mind, Zega Avseck, says this level of granularity is vital for the next generation of drug discovery. Most modern treatments target specific proteins within specific cells, so knowing the exact cellular context is crucial. Alpha Genome2 can reportedly identify twice as many gene expression variants as previous models. There is a trade-off, however, in terms of computational cost. This model is significantly more expensive to run than the original alpha genome. But DeepMind argues the ability to accurately simulate biological systems without the need for constant physical retraining represents a major shift toward what they are calling virtual cells. [08:51] Thatcher Collins: It is a potent reminder that while generative AI for consumers is currently under heavy legal fire, the specialized scientific applications are accelerating. This contrast is reflected in the financial markets as well. Bloomberg is reporting that Anthropic's initial public offering is roughly six times oversubscribed. Despite the broader regulatory climate, investors are valuing the company at up to one point four trillion dollars. Shares are expected to begin trading on the NASDAQ under the ticker symbol ANT on November19 th. It suggests that the market sees immense value in Anthropic's safety first branding and its potential for long-term institutional growth. [09:30] Nina Park: Despite the ongoing litigation risks we discussed, investor confidence in the creator of the Claude models seems largely unshaken. At the same time, we are seeing the emergence of new infrastructure startups aimed at making these models safe for enterprise use. A company called Trusted Router launched three encrypted fusion models this week. These systems combine multiple large language models, including those from Meta and Envidia into a single interface while keeping user prompts fully encrypted. The data remains encrypted until it reaches the provider's verified secure hardware, ensuring that even the model providers cannot see the raw sensitive data being processed during inference. [10:12] Thatcher Collins: The founder of Trusted Router, Joseph Perla says he is aiming to build what he calls the New York Token Exchange, or NYTE. The goal is to turn AI inference into a competitive commodity market, where businesses can trade AI capacity, much like they trade electricity or natural gas. Their flagship models, named Prometheus and Liberty, use a panel of independent AI judges to evaluate and synthesize answers from multiple sources. They claim this multi-model approach can save businesses up to sixty percent on token costs compared to relying on a single high-end model like the upcoming Claude Fable, all while maintaining higher accuracy and lower latency. Fatcher, [10:54] Nina Park: listening to all of this, it really feels like the industry is splitting into two parallel tracks. In one lane, we have these incredible scientific breakthroughs in genomics and sophisticated market exchanges for compute. In the other lane, there is a massive and growing wall of legal and social resistance. Open AI's research into fuzzy robustness is a clear attempt to close that gap by providing technical proof of safety, but the current wave of lawsuits suggests that purely technical solutions might not be enough to satisfy the public's demand for transparency and corporate accountability. [11:26] Thatcher Collins: The friction we are seeing isn't just about whether the models work or how fast they can process tokens. It is about how they are governed. Whether it's an autonomous dot making a purchase on a credit card or a chat bot integrated into a middle schooler's backpack, the demand for institutional guardrails is louder than it has ever been. We're moving out of the era where the primary question was, what can this model do? We have entered the era of who is responsible when it fails. That shift in perspective is going to dominate the boardroom and the courtroom for the foreseeable future. [11:58] Nina Park: It is a transformation that will undoubtedly define the next few years of AI development and regulation. We will be watching very closely as Anthropic officially hits the public market next week, and as the California courts begin to weigh the merits of these public nuisance claims. These are the test cases that will set the precedent for the entire frontier AI category as it matures into a regulated industry. [12:25] Thatcher Collins: Indeed. The technical and legal landscapes are finally converging after years of running in parallel, and the stakes for the labs involved really could not be higher. We are seeing the true birth of the AI economy, growing pains in all. It is a transformation of labor and science that is only just beginning. [12:44] Nina Park: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. You can find our research notes and further coverage of these stories at MB. NeuralNewscast.com. Neural Newscast is AI assisted, human reviewed. View our AI transparency policy at Neuralnewscast.com. [13:04] Announcer: This has been model behavior on Neural Newscast. Examining the systems behind the story.