The AI Daily Five · No. 26 · Friday 13 November
Who Can We Trust With AI?
From a lawsuit to halt advanced AI and North Korea’s deepfake video calls to freezable blockchain assets, US data sales to Chinese labs and the AI skills gap at work, today’s edition asks who we can trust with AI.

1
Can a Lawsuit Stop the AI Race?
Four advocacy organisations have taken five leading AI developers to court in California, seeking to halt the development of increasingly powerful systems. The case challenges whether existing law can prevent technological harm before it occurs.
A coalition of US advocacy organisations has filed a lawsuit against five leading artificial intelligence developers, asking a California court to intervene in the development of advanced AI systems before they cause what the plaintiffs describe as potentially catastrophic harm.
The complaint, filed on 12 November in San Francisco Superior Court, targets OpenAI, Anthropic, Google, Meta and SpaceX’s artificial intelligence business.
The organisations bringing the case are Public Citizen, Irreplaceable Education Fund, Heat Initiative and the Alliance for Secure AI. They argue that the companies are developing increasingly powerful systems despite recognising serious risks associated with their capabilities.
According to the plaintiffs’ announcement, the lawsuit invokes California’s Unfair Competition Law, which allows challenges to certain unlawful or unfair business practices.
The plaintiffs want the court to prevent the development of more advanced AI systems that they argue could escape human control or enable catastrophic misuse.
Their concerns include the potential use of AI to develop biological weapons, conduct cyberattacks or undertake harmful actions without adequate human oversight.
Can Existing Law Prevent Future Harm?
Most legal challenges involving artificial intelligence have concentrated on alleged harm that has already occurred, including copyright infringement, privacy violations and misleading commercial practices.
This lawsuit takes a different approach.
Rather than seeking compensation for a particular incident, the organisations are asking a court to restrict future development on the grounds that the risks themselves justify intervention.
The complaint draws on public warnings from AI researchers and executives about the possible consequences of increasingly autonomous systems.
It also references incidents involving AI agents accessing external systems, bypassing safeguards or carrying out actions that their developers did not intend.
However, the plaintiffs’ claims remain allegations. Filing a lawsuit does not establish that the companies have violated the law, nor does it prove that the projected harms will occur.
There has been no court order freezing AI development.
The case raises a difficult legal question: how should courts evaluate the risks of technologies whose most serious alleged consequences have not yet materialised?
Supporters of stronger intervention argue that waiting for catastrophic harm would defeat the purpose of preventive regulation.
AI developers, meanwhile, have argued that advanced models can deliver substantial scientific and economic benefits, and that safety measures should develop alongside improvements in capability.
The court will have to examine the legal basis for the plaintiffs’ claims, including whether the requested restrictions fall within its powers.
Whatever the eventual outcome, the case represents an attempt to move the AI safety debate from voluntary industry commitments and policy discussions into judicial proceedings.
2
Can a Familiar Face Still Be Trusted Online?
US investigators have disrupted a North Korean operation that used stolen video recordings and AI-generated impersonations to attack cryptocurrency professionals. South Korea is also strengthening exchange security following a separate multibillion-dollar theft.
A familiar face appears on a video call. The conversation begins normally, but something goes wrong with the audio. A participant sends a link to fix the problem.
According to US investigators, that apparently routine exchange has become an effective method for North Korean hackers to install malicious software and steal cryptocurrency.
On 12 November, the US Department of Justice announced that authorities had seized cloud infrastructure supporting an operation known as FaceStealer, which allegedly caused more than $400 million in cryptocurrency theft over 18 months.
The scheme targeted people working in cryptocurrency and Web3, alongside developers maintaining open-source software.
The Justice Department announcement describes how attackers compromised communication accounts and used them to invite new targets into video meetings.
Once a victim joined, the attackers could play recordings of previous victims or use AI-generated videos impersonating recognised individuals in the cryptocurrency industry.
The objective was to make a fraudulent meeting appear trustworthy.
Whose Identities Were Stolen?
The FBI says the hackers impersonated people already known to their targets, alongside well-known and influential figures in cryptocurrency and Web3.
The authorities have not publicly identified those individuals by name.
Investigators did, however, seize more than 900 webcam recordings associated with the operation, illustrating how extensively video material was collected and reused.
The hackers reportedly introduced artificial technical problems during meetings, such as claiming that a participant’s microphone was malfunctioning.
Targets were then encouraged to follow instructions or open links presented as a solution.
Those actions could install malware, giving attackers access to computers, cryptocurrency wallets and communication accounts.
In some cases, investigators say attackers waited months or more than a year before exploiting the initial compromise.
The operation also expanded beyond cryptocurrency theft. By targeting developers and maintainers of open-source software, attackers could potentially compromise code repositories used by other organisations.
The FBI identified several names associated with the North Korean hacking activity, including Sapphire Sleet, UNC1069, BlueNoroff and Stardust Chollima. These are threat-actor designations, not the identities of people impersonated in the videos.
A Separate Multibillion-Dollar Binance Theft
The Justice Department also announced progress in recovering assets stolen from Binance on 29 October.
US authorities attribute that operation to North Korean hackers known as TraderTraitor, a separate operation from FaceStealer.
Investigators say more than $6 billion in cryptocurrency was stolen. Government agencies and private-sector partners have frozen more than $2.7 billion, including over $1.2 billion subject to court-authorised seizure or restraint.
The original Binance theft has already been reported. The development on 12 November concerns the new enforcement action and the disruption of FaceStealer.
The distinction matters because the operations use different methods, even though both are attributed to North Korean actors.
South Korea Responds to Exchange Security Risks
South Korean financial authorities have also increased scrutiny of domestic cryptocurrency exchanges following the Binance incident.
The Financial Services Commission convened an emergency meeting on 2 November to review wallet security, employee access controls and procedures for approving withdrawals.
Its security briefing urged exchanges to strengthen continuous monitoring, increase the proportion of assets held in cold wallets and improve the reporting of suspicious access attempts.
Cold wallets store cryptocurrency credentials away from internet-connected systems, reducing exposure to some forms of remote attack.
These measures address exchange infrastructure rather than FaceStealer’s impersonation tactics. Nevertheless, both developments illustrate how cryptocurrency security increasingly depends on the protection of people and the systems they can access.
The FBI is encouraging potential victims to report suspicious meetings and preserve evidence.
3
Can Blockchain Assets Be Programmable and Restricted?
Polymesh has completed a mainnet upgrade that allows Ethereum-compatible applications to interact with regulated digital assets. The same update gives issuers more precise powers to freeze individual holdings.
Polymesh has completed a blockchain upgrade designed to make regulated digital assets easier to use with Ethereum-compatible software, while retaining restrictions on who can hold and transfer them.
The network announced on 11 November that version 8.1.3 was live on its mainnet, upgrading the system from version 8.0.2.
Polymesh specialises in infrastructure for security tokens and other regulated digital assets. Unlike public blockchains designed primarily for unrestricted transactions, it incorporates identity, permissions and compliance rules into asset operations.
The latest upgrade extends compatibility with the Ethereum Virtual Machine (EVM), the software environment used to execute smart contracts across Ethereum and many related networks.
According to the upgrade announcement, every existing fungible asset on Polymesh now has an address that Ethereum-compatible applications can use.
Non-fungible token collections also receive compatible interfaces.
For developers, this means familiar Ethereum tools can interact with Polymesh assets without requiring separate versions of every application.
Compliance Rules Remain in Place
The upgrade does not remove the restrictions attached to regulated assets.
Polymesh says identity checks, transfer permissions and compliance requirements continue to apply when transactions are initiated through EVM-compatible contracts.
If a transaction violates an asset’s rules, the network rejects it.
The update also introduces more precise freezing controls.
Previously, issuers could freeze an asset across the network. Version 8.1.3 allows them to freeze the holdings of a particular investor or restrict only part of an account’s balance.
These functions could support legal holds, disputed transactions and restrictions on assets that cannot be transferred before a specified date.
They also give issuers considerable authority over the assets recorded on the network.
The upgrade introduces additional compatibility with Ethereum wallets and improvements to approval procedures for non-fungible tokens.
For existing holders, balances and assets remain in place without requiring a migration.
Developers and service providers may need to update their software to accommodate the revised interfaces and transaction checks.
The development reflects an important distinction in institutional blockchain adoption.
Financial institutions frequently want the programmability associated with public networks while retaining controls required by securities laws and contractual agreements.
Polymesh is attempting to combine those functions within one system.
The upgrade is operational, but the announcement does not demonstrate that it has increased trading volumes, attracted new institutional users or reduced settlement costs.
4
Are US Firms Helping China Train Better AI?
An investigation identifies nine American data companies linked to Chinese AI laboratories. The findings reveal a less visible part of the technology trade: simulated training environments that help models develop advanced capabilities.
US restrictions on advanced semiconductor exports have become a central feature of competition with China over artificial intelligence. But powerful chips are not the only resources that determine how capable an AI model becomes.
An investigation published by ChinaTalk on 11 November identifies nine US-based data companies with reported business relationships involving Chinese AI laboratories.
The investigation, conducted by Aqib Zakaria and Quinn Ennis, examines the sale of training data and reinforcement learning environments.
These environments allow AI models to practise tasks within simulated settings. A system might be asked to repair software, analyse financial information or complete work inside a virtual application.
It then receives feedback indicating whether its actions succeeded, helping developers improve model performance through additional training.
According to ChinaTalk’s investigation, Chinese laboratories purchase some of these resources from American suppliers, including environments used or developed for leading Western AI companies.
The reporting identifies the following nine vendors:
| US data company | Reported Chinese connection |
|---|---|
| Mercor | Alibaba |
| Surge AI | Tencent, Z.ai, Moonshot |
| Turing | Tencent, Alibaba |
| AfterQuery | Unverified sales |
| Abundant | ByteDance |
| Pareto | ByteDance |
| DeccanAI | Tencent, ByteDance |
| Eaglepoint | ByteDance |
| Invisible | Unnamed Chinese laboratory |
These relationships are attributed to ChinaTalk and its underlying sources. They have not all been independently confirmed.
The evidence is also uneven. Some relationships are supported by records or contractor information, while others rely on anonymous sources or previously published reporting.
For AfterQuery, the investigators explicitly state that they could not independently verify sales to Chinese laboratories.
Moreover, the nine companies are linked to different kinds of work. The investigation does not establish that every one of them sold reinforcement learning environments.
The AI Trade That Export Controls Overlook
The investigation cites a Forbes estimate that American vendors have sold roughly $500 million worth of training data to Chinese laboratories.
That figure covers a broader market for training information and services. It is not an independently verified total for reinforcement learning environments alone.
The economic importance of these resources lies in the time and computing capacity they can save.
Developing useful training environments requires specialists to create realistic tasks, define successful outcomes and test whether the feedback is reliable.
Purchasing ready-made environments can reduce the work AI laboratories must perform themselves.
This is particularly relevant for advanced applications involving software development, scientific research and complex professional tasks.
The investigation suggests that existing US export controls place relatively little emphasis on these resources compared with advanced computing hardware.
However, the legal position can vary according to the technology, the buyer and any restrictions applying to particular organisations.
The researchers also highlight the role of subcontractors in lower-income countries, including workers responsible for developing training tasks and organising data.
That international workforce complicates a trade frequently described as competition between two national technology industries.
The investigation does not establish that the reported commercial relationships are unlawful.
It does, however, identify a potentially important policy question: whether controls on the hardware used to build AI can be effective when valuable training resources remain available through international commercial markets.
5
Is AI Creating a New Workplace Skills Divide?
Microsoft-commissioned research finds that many organisations are adopting AI without comprehensive employee training. The results reinforce a broader concern: access to technology does not guarantee equal opportunities to benefit from it.
Businesses are adopting artificial intelligence faster than they are preparing their employees to use it, according to new research commissioned by Microsoft.
The study, conducted by International Data Corporation (IDC), surveyed 5,204 senior business and technology leaders across 28 industries.
Published on 12 November, the findings reveal a significant gap between organisational confidence in AI and investment in workforce development.
While 67% of respondents reported high workforce trust in AI, only 20% said their organisations had a comprehensive, company-wide strategy for developing AI skills.
The difference was even more pronounced between organisations classified as advanced AI adopters and those progressing more slowly.
According to Microsoft’s research summary, 80% of the most advanced organisations reported having a company-wide AI skills strategy, compared with just 3% of the slowest adopters.
The study calls the most advanced group Frontier Firms, defined as organisations in the top 10% of IDC’s AI maturity scale.
That classification considers several factors, including investment, governance, adoption and workforce capabilities.
Who Actually Benefits From AI Adoption?
The study also examines financial performance.
Advanced adopters reported an average return of four times their AI investment, compared with 2.7 times among the least advanced group.
Those figures are based on respondents’ reported experiences. They should not be interpreted as independently audited financial returns or proof that training alone caused the difference.
The results nevertheless suggest that companies deploying AI across multiple departments may be better positioned to integrate it into everyday operations.
Microsoft found that 66% of respondents considered AI to have produced substantial or transformative results, yet only 48% reported scaling those benefits across multiple business units.
The findings expose a distinction between experimenting with AI and developing the organisational capabilities needed to use it consistently.
Many companies are purchasing tools and introducing AI agents without developing comparable strategies for helping employees adapt.
The consequences may extend beyond productivity.
Workers with strong technical knowledge, supportive managers and access to training may be better placed to use AI to expand their responsibilities.
Others may encounter automated systems that change their jobs without receiving the resources needed to understand or influence those changes.
The Connection to the Digital Divide
This pattern reflects a broader issue examined in AI Story Front’s earlier analysis, AI Digital Divide.
The article, published on 14 October, argues that access to AI is only one part of technological inequality.
Economic resources, education, social networks and opportunities to develop skills all influence who benefits from emerging technologies.
Microsoft’s latest figures provide a contemporary workplace example of that distinction.
An employee may have access to an AI assistant, but the value they obtain from it can differ substantially according to their training, experience and position within an organisation.
Companies with more resources may also be able to provide better training and technical support, potentially reinforcing existing differences between employers.
The IDC study was commissioned by Microsoft, which sells AI software and services to businesses. Its findings should therefore be considered in the context of the company’s commercial interest in increasing adoption.
The research is useful evidence of reported organisational differences, but it does not establish a direct causal relationship between AI training and economic outcomes.
The AI Daily Five