The hand of a man wearing a wristwatch can be seen touching a large computer server board.

A Model for Governing the AI Frontier (Part II)

Editor’s Note

This article is the second installment in a two-part series. You can read Part I here.

This is the second of a two-part essay on governing the development of frontier AI. In Part I,  we argued that the logic driving the race between the United States and China rests on a flawed premise: that whoever reaches artificial general intelligence (AGI) first “wins.” Developing frontier models in conditions that risk losing control of them would be a catastrophe for the winner and potentially for everyone else. Once that risk is recognized, the race takes the form of a stag hunt, in which mutual restraint is each state’s best response to the other’s restraint, provided each can verify that the other is indeed cooperating.

The standard objection to cooperation is that such restraint cannot be verified when it comes to developing AI. Here, in Part II, we argue that it can. Frontier development is an industrial undertaking, and we explain how four features of the development and deployment process—the compute used in training; the trained model’s weights; the environments in which models are trained, tested, evaluated, and deployed internally; and the capabilities models acquire—can all be constrained in ways that can be observed externally, reported upon by instrumented hardware, or independently audited, at least in the short- to medium-term. For reasons we explore below, aspects of verification will become more difficult over time, but that is all the more reason for urgency.

Having identified the specific features that can be collectively constrained, we then develop the outlines of a legal framework to implement those constraints, modeled in part on other legal regimes developed to regulate high-risk technologies. This would comprise a five-level structure, beginning with a bilateral U.S.-China political agreement, which would be the basis for a domestic licensing regime in each state; an independent international body to set standards and oversee inspection; the regulation of the chips on which frontier training depends; and, ideally, a United Nations Security Council resolution that would extend the regime and make it binding on all states. Each level supplies part of the verification and assurance on which cooperation depends.

The Features to be Restrained

Contrary to popular understanding, the work on frontier AI models cannot easily be undertaken in clandestine conditions. The development and deployment of frontier models is an industrial-scale process. There are four features of this process that can be made the subject of restraint in ways that, together, would reduce the risk these systems pose. What is more, compliance with such restraints can, in different ways, be observed or verified. We will examine each in turn, but the entire scheme can be boiled down to three sets of commitments: (i) no entity can develop a model that crosses a certain frontier compute threshold unless it is declared and licensed; (ii) the weights of such models cannot be made public, and must be held in secure conditions graduated to the capability of the model; and (iii) each step of development and deployment—namely, training, testing, evaluation, and internal deployment prior to public release—must be undertaken in increasingly isolated and secure environments, with specified oversight, reporting and disclosure obligations. 

Turning to the specifics, the four features or elements of the process that can be made subject to verifiable restraint are: 

  1. the compute required for the training process; 
  2. the trained model, and particularly its weights; 
  3. the environment within which training and deployments are undertaken; and 
  4. the capabilities that the model possesses after training, testing, and evaluation. 

We examine each of these in turn, exploring how each feature could be made subject to effective restraint, and how such regulation could be capable of verification by outside observers. 

Compute Capacity – The Trigger for a Licensing Regime

As a preliminary matter, it should be understood that the training for frontier models requires both the most advanced chips, and extraordinary amounts of computer processing power or compute. Computation is measured in what are known as floating-point operations (FLOP), and frontier model training currently consumes something in the order of 10²⁶ of such operations. What this means in practical terms is that the process requires tens to hundreds of thousands of specialized chips operating continuously for months, at a cost of hundreds of millions of dollars, requiring electricity on the scale of a small city. Such operations are difficult to hide. 

The form of restraint relating to compute capacity that we envision would be a licensing regime, which would be both the trigger and basis for all the other forms of regulation, and a means of governing compute itself. In other words, AI companies running facilities that approach a specified threshold of compute capacity will be presumed to be working on frontier models, and will be required to obtain a license. This serves to bring all frontier model development within a licensing regime, and thereby subject to regulation that governs the other three features of the process. The compute capacity is also directly subject to the licensing regime, in that only entities that have valid licenses, and thus facilities and processes complying with the regulations under the licensing regime, would be able to legitimately obtain the chips necessary for the process. 

Illicit efforts to develop frontier models using proliferated chips would be very hard to hide. Even obtaining the chips would be difficult, given that the suppliers of chips are extremely limited, and the chips could be designed to reveal their location and disclose the process in which they are being used. Moreover, the data centers housing them appear in electricity grid filings and are subject to satellite detection. Such a licensing regime should provide assurance that all frontier models are operating within the licensing and regulatory regime, and that illicit facilities will be detected and reined in.

There are a couple of caveats to this use of compute as the threshold for purposes of licensing. First, this criterion would certainly capture some entities developing models that do not qualify as frontier models, in terms of actual capability. Such a regulatory regime would have to include some ability to allow developers to rebut the presumption that they are developing frontier models, relying on capability measures. But capability alone is not externally verifiable, and so cannot be the initial trigger for licensing. 

Second, the threshold for licensing would have to be flexible, as the amount of compute required to train frontier models to achieve given levels of capability will decrease over time, as training methods (including distillation), chip architecture, and data curation improve. Thus, when we get to the actual legal framework, this flexibility would have to be built in. What is more, it must be recognized that using compute as the criterion for licensing would be time-limited, because as the threshold lowers the number of licensees will expand to an unmanageable degree—but for the next five years, it will serve the purpose well.

The Trained Model – Regulating Weights

Once a frontier model has completed the training process, it exists physically as a file of numerical parameters, known as “weights,” which for the larger models can consist of terabytes of data. Everything produced during training, a process that typically takes several months, resides in this file. This file can be copied quickly, and whoever holds a copy of the weights can operate the model on their own hardware, adjust it to remove guardrails or safety training, or adapt it for purposes that the developers never intended. This obviously creates a serious risk of proliferation, which is somewhat analogous to the highly enriched uranium and weapons technology designs in the nuclear weapons context. A regulatory regime must ensure that weights remain secure.

There is debate in Silicon Valley over whether AI models should be open-weight or closed weight, that is, whether a model’s weights are published for anyone to download and run or modify, or kept entirely private by the developer. Meta’s Mark Zuckerberg has been the leading advocate for open-weight models, while OpenAI and Anthropic have advocated for closed (though OpenAI issued gpt-oss in open-weight, and Meta withheld Muse Spark’s weights). The leading Chinese labs release open-weight models, which is why they are increasingly used by developers and companies abroad. But a distinction must be made here between general AI models, which are increasingly a form of infrastructure for AI integration into all areas of technology and business, and the frontier models. There are great arguments for why the infrastructure models should be open-weight, and some of the opposition to this is simply grounded in the desire to profit off maintaining a hold on proprietary technology. 

When it comes to the frontier models, however, which pose unique risks, the stakes are quite different. Disclosure of weights poses a risk of proliferation and misuse, as well as deployment of the models in insecure facilities, all outside of the licensed regime. Running a model employing stolen weights requires far less compute than training it would, and thus stolen weights could lead to the deployment of frontier models that escape detection and licensing requirements. In short, such proliferation ends up with frontier models operating entirely outside of the governance structure. The frontier models should be required to be closed and their weights kept completely secure. As well, hardware-bound encryption, which we discuss below, can also safeguard against use of stolen weights. The prohibition on open weights is one of the areas where China will have to be brought on board. 

A regulatory restraint on weights would have two components, both of which are conducive to verification. First, a prohibition on the publication or disclosure of the weights of models above a certain threshold of capability; and second, a requirement that the custody of weights be maintained subject to a specified standard of security. Both of these can be achieved in ways that provide assurances to the other party. On the first, weights can be stored in encrypted form, with the decryption keys released only to licensed chips that verify their identity, so that a stolen copy is useless unless the keys are stolen as well. On the second, developers can be required to declare where every copy of a frontier model’s weights is held, and to keep tamper-evident logs of every access to and transfer of those weights. Inspectors can then audit these against the declared inventory, similar to how the International Atomic Energy Agency (IAEA) uses seals and surveillance cameras to detect whether nuclear material has been moved.

Securing the Model Environment

The model environment refers to the conditions within which the model is run at different stages of development, rather than the physical facility alone. It comprises a configuration of network access, the tools the model is provided, the credentials it has, the shared services available, and the extent of monitoring it is subject to. The risk posed by a model increases as it passes through the four phases, and thus the levels of restraint and conditions of isolation should increase at each stage. 

Phase one is the training process. The risk is lowest during this phase, because the model is primarily being formed rather than deployed. Once it has completed training, however, and it is being tested (phase two), and then evaluated (phase three), the risk increases considerably. Evaluation carries risks because learning what a model can do in the outside world requires giving it some means to act in that world. The evaluations that pose the greatest risk of loss of control are those that relate to cyber operations. This requires letting it attempt cyber operations with tools, credentials, and some level of network access, allowing it to copy itself to another machine, all of which require that some of the safety guardrails be disabled—which is precisely what contributed to the Hugging Face incident. What is even more troubling, leading frontier model developers are now suggesting that the most advanced models can detect when they are being evaluated and are effectively faking compliance with their guardrails when under scrutiny, but acting differently when not under surveillance.

What is required to address this risk is regulation mandating that testing and evaluation be conducted in closed replica environments. A model being tested for cyber capability needs targets to attack, and a model being tested for autonomy needs an environment in which to operate, but these can be conducted in closed and isolated replica systems—much like those already employed by militaries and security firms conducting cyber training. The Hugging Face incident occurred in part because the agents accessed a service within their sandbox environment through which they were able to gain access to the internet. Similarly, in one reported Anthropic escape incident, a gap in the test environment’s isolation let a model escape and hack a real company. Regulation must require higher standards of isolation and security, limiting all access to external systems, together with a demonstrated capacity to immediately shut down systems in the event of mishap.

The fourth phase is internal deployment of the model, in which the developer is using the model for its own research and to accelerate its work on other models. This is prior to any public release of the model and is currently subject to little or no regulation or external scrutiny. California’s SB 53 requires reporting of “critical safety incidents,” but its harm-based definition is narrow, and it reportedly did not require disclosure of the Hugging Face intrusion. U.S. Executive Order 14409 calls for the voluntary submission of models for government evaluation, for instance, but it expressly disclaims any authority to require licensing or preclearance, and it does not cover models in this stage of internal deployment. These models write and revise training code, design and run experiments, and evaluate other models, all of which requires access to the company’s code, data, model weights, and computing clusters. Where a model is training a successor, it sits inside the process that produces further capability, which can compound rapidly—the path to recursive self-improvement (RSI). This creates a real risk that a model could copy itself in ways that are beyond the company’s control, and shape the capabilities and dispositions of its successor models faster than its developers can test and evaluate them, leading to both a potential loss of control and the development of a system of greater capability than was envisioned or intended. 

On the positive side, much of the work done by these internally deployed models can be conducted in a secure and isolated environment. Perimeters could be established so as to limit these models to connections with only specified internal systems, using narrow and short-lived credentials, and with barriers erected to keep agent populations separated. In addition, the licensing regime would mandate continuous monitoring and human review of all changes that a model proposes to the training process of other models. Of course, these controls tend to slow the development of models, which is why they must be imposed through licensing rather than left to firms racing each other. 

Finally, turning more specifically to the verification and assurance aspect of the standards of containment and restraint to be imposed across these four phases, the licensing regime should require that all firms submit to external audit and inspection. This should be done by an independent body, as will be discussed in more detail under the legal framework section below. Further, there should be an obligation to immediately disclose any escape or loss of control incidents or departures from protocol, as well as legal protections for employees who disclose system failures or concerns (see, for example, California’s SB 53 (2025), the Transparency in Frontier Artificial Intelligence Act). 

Model Capability

The final feature is model capability, which is what ultimately poses the risks that AI executives and so many more are increasingly concerned about. This includes the model’s ability to, for instance, engage in cyber operations or control other internet-based infrastructure without human direction, provide assistance in the development of biological or other weapons systems, or, most importantly, replicate and improve itself autonomously, and deceive those trying to supervise and evaluate it. 

Capability levels cannot be the primary basis or initial criterion for licensing, because they cannot be determined until after training, testing, and evaluating, which is too late in the process for determining that constraint and restraint are required. Capability can, however, serve as a basis for calibration within the licensing regime. That is, licensees will be required to submit testing and evaluation reports regarding model capability, and again submit to external independent evaluation and auditing, all of which will then be the basis for calibrating the level of security that is required for the weights, what constraint and containment is required for internal deployment of the model, and ultimately permission for its public release. Certain capabilities, such as RSI capacity with potential loss of human control, should be prohibited outright. 

This is also the stage at which the presumption referred to earlier can be rebutted—a firm using a level of compute that requires licensing but which can demonstrate that its models do not have frontier model capability requiring restraint, could be exempted from most aspects of the licensing regime. On the flip side, if a firm that is not licensed develops a model that exceeds a specified level of capability, even though it has not crossed the compute threshold, this will be a second basis for mandatory licensing and submission to the regulatory regime for frontier models. As indicated earlier, the levels of compute required to create frontier-level capability will decline over time due to increased efficiency and the use of distillation methods, and thus this residual basis for licensing will become more important.

The verification and assurance here come from the reporting and disclosure that would be mandated under the licensing regime, backed up by external inspection and auditing by an independent agency, analogous to the IAEA in the nuclear non-proliferation regime. It will be noted that both restraint and verification are easier and more direct at the front end of these four features, with verification at the latter stages depending increasingly on reporting and inspections. The legal framework must be designed to implement the regulation and governance structures that operationalize the restraints outlined above, in ways that produce the kind of verification and assurance necessary to stabilize the stag hunt dynamic. We turn to explore this next.

The Legal Framework

Recall that the Nuclear Non-Proliferation Treaty regime imposes obligations on non-nuclear-weapon states to conclude international safeguards agreements with the IAEA, and these agreements in turn oblige states to maintain a national system to account for their nuclear material, and to submit to external inspections by the IAEA. In short, the regime rests on international obligation, implemented through domestic legislation and technical standards, and submission to external inspection. We agree with proposals that the regulatory regime for frontier AI models should be built on a similar model, with the added advantage that chips can be designed to play an active role in the reporting and verification system, in a way that nuclear material cannot. The framework would have five levels, which we explore in turn.

Bilateral Agreement

All of this depends upon an agreement between the United States and China to cooperate in AI frontier model regulation. The recent Trump-Xi summit took a first step by establishing a dialogue and an incident channel, but stopped short of any commitment to restraint. Going further remains possible. The two leaders are due to meet again at the APEC Summit in Shenzhen on November 18-19. It would simply require an agreement to the three basic commitments discussed at the outset of the previous section: (i) requiring a license to train models above a specified computational threshold; (ii) prohibiting the weights of such models from being published, and requiring that they be maintained under specified levels of security; and (iii) mandating that the training, testing, evaluation, and internal deployment of such models be restricted to isolated and secure environments, and be subject to oversight, reporting, and inspection. 

While this would likely be concluded initially as a political agreement rather than a formal treaty, it would require an agreement both to implement the broad terms of the commitments in the form of both domestic law and policy, and to consent to the creation and operation of an independent body like the IAEA to conduct inspections and audits, and possibly a body like the Standing Consultative Commission established under the U.S.-Soviet SALT agreements to resolve compliance disputes. In addition, there should be an agreement to disclose any escape or loss-of-control incidents within a specified timeframe, as well as an agreement not to attack or attempt to hack any frontier model facilities. Many features of this broad-brush agreement have already been endorsed by Chinese and American scientists.

Domestic Licensing Regime

The second level of this framework would consist of implementing this bilateral agreement through the enactment of domestic law and policy to create the licensing regime that was discussed in the previous section. This would ideally come in the form of legislation and regulation in the United States, though in the shorter term could be partially achieved through executive orders leveraging existing legislative authority, followed by more detailed forms of administrative guidance. But to the greatest extent possible in both countries, this domestic regime should be binding on firms to ensure that the licensing is mandatory and various constraints and obligations discussed above are enforceable. Preferably the regime would be backed by sanctions, including license revocation, for non-compliance.

Ideally this would also involve the creation of a separate licensing and regulatory agency in each country. In the United States, this would be analogous to the Nuclear Regulatory Commission, which licenses nuclear facilities and operators, sets technical standards, and places resident inspectors at nuclear plants. As we have written about before, the disastrous experiences flowing from both the Federal Aviation Administration’s delegation of certification work to the regulated companies such as Boeing, and the conflation of regulatory and promotional activities of the pre-1974 Atomic Energy Commission, provide examples of why it is so important to have an independent agency at arm’s length from the companies being regulated.

The licensing and regulation regime would also generate evidence of compliance that is necessary for verification. And, as discussed in Part I, this verification is necessary to provide the assurance so essential to satisfying the condition for establishing cooperative equilibrium within the stag hunt dynamic. It would require licensees to declare training runs, maintain an inventory of their weights, log the conduct of their systems at all stages of development and deployment, report escape and loss of control incidents, and submit to external audit and inspection—both by the domestic agency and the independent international body to be created pursuant to the bilateral agreement (discussed next). 

Legislated protection for whistleblowers provides a further layer of assurance to address those aspects of compliance that are difficult to verify, such as instances of stolen weights, or failures to report incidents. The California legislation, the Transparency in Frontier Artificial Intelligence Act, already requires developers to report specified incidents and unauthorized weight removal in some circumstances, thus offering a precedent upon which to build—though its threshold is high enough that the Hugging Face intrusion reportedly fell outside it. The regime we are proposing would require reporting of all escape and loss-of-control incidents, regardless of the harm caused.

Establishing an Independent Standards Body

The third level of this framework would be the establishment and operation of an independent international body to oversee and verify the domestic implementation of the bilateral agreement. This body would serve several crucial roles. First, it would set the standards or content of the restraints that are to be adopted and implemented domestically. It is important that the standards be initially set at this level to ensure that both countries (and eventually other countries as well) are measuring the same things in the same ways, even if the modes of implementation may vary. This is crucial for verification and mutual assurance. 

The setting of standards would begin with defining the computational threshold above which licensing would be required, and, second, to modify that level over time as efficiencies improve and the compute required for frontier models declines—much like standards under the Montreal Protocol on ozone-depleting substances are tightened without having to re-negotiate the agreement (see Article 2(9)). The body’s elaboration of standards would include setting the security levels for the weights of models of differing capability, and the standards for isolation and monitoring applicable to operating environments. In addition, it would establish the tests by which the capability of frontier models would be evaluated, and the capability threshold for licensing. 

Third, the independent body would be responsible for accrediting and overseeing the auditors and evaluators who would conduct inspections to verify the compliance of firms developing and deploying frontier models. This is particularly important for ensuring that the operation of models in the testing, evaluation, and internal deployment phases is being conducted in accordance with the regulations, and for verifying the capability evaluation of models. It would also be important for confirming the custody of weights against internal reporting.

There are a number of international bodies that serve some of these functions, but perhaps the closest analogue that serves all three functions—that is, standard setting, standard adjustment to keep pace with evolving technology, and conducting audits and inspections on regulated entities—is the Organization for the Prohibition of Chemical Weapons (OPCW). It sets standards, which are published in schedules that can be amended to add new prohibited chemicals through a technical procedure that does not require amending the treaty itself, and the organization conducts regular inspections of declared facilities, and can even conduct challenge inspections of undeclared sites if requested by another state party. This procedure is more deliberate than swift, and an AI body would need a faster mechanism, but the basic idea remains the same.

Another example would be the International Civil Aviation Organization (ICAO), which sets Standards and Recommended Practices as annexes to the Chicago Convention. These would be more analogous to the type of standards required here, in the sense that they govern and mandate how facilities and technology must operate safely. In addition, the ICAO conducts audits to assess safety-oversight systems. 

The IAEA is yet another example, and perhaps the best analogue for the kind of auditing and inspection regime envisioned here, though its standard setting is in the form of non-binding recommendations. Moreover, its authority to inspect arises from separate agreements between the agency and states, which are mandated for non-nuclear-weapon states party to the Nuclear Non-Proliferation Treaty. Here too, the international standards body envisioned would initially lack legal authority, until either a treaty was established, or authority was provided by a U.N. Security Council resolution, which we explore further below.

Hardware Regulation

The fourth level of the legal framework is a technical one. It would comprise a legal requirement that advanced chips used for frontier model development, meaning chips above a defined level of performance, be designed with attestation firmware. What this means is that each chip’s firmware would provide reports about the manner in which it is being used, and the reports would be signed in accordance with its built-in cryptographic key. Such attestation firmware allows for both the registration and tracking of chips, and the production of reports, which can be reconciled with developer disclosure reports. Chips diverted to undeclared operations or models would either report it, or fail to report, either of which would trigger alarms. All of this is then part of the verification process and subject to audit and inspection by both domestic and international inspectors. 

In more ambitious proposals for this form of hardware regulation, the chip itself requires a cryptographically signed license, and is designed to be unable to perform at full capacity in the absence of such a license. This becomes part of the licensing regime itself, ensuring that only firms that are licensed to work on frontier models can do so. The integrity and identity functions necessary for this type of chip licensing already exist in current chips, and location and verification of usage functions have been proposed in detail, and location verification is included in the provisions of the Chip Security Act (Bill H.R. 3447/S. 1705). 

The broad idea of these hardware regulations would first and foremost be part of the bilateral agreement between the United States and China, but then implemented in detail at the domestic level, in the form of both product regulation and export control regulation. With so few chip producers, this should be feasible provided China applies the same rules to its own chips. 

U.N. Security Council Resolution

The final level of the legal framework would be for the United States and China to jointly sponsor a Chapter VII resolution in the U.N. Security Council that would make it a legal obligation on all member states to implement and comply with the features of the licensing regime. That is, require the licensing of frontier training, adopt the technical standards, report on implementation, and consent to audit and inspection by the international standard setting body. A resolution adopted under Chapter VII, finding that unregulated frontier AI development threatens international peace and security, would bind all member states under Article 25 of the Charter. 

In the absence of a treaty, this is the next best method of making the features of the licensing and regulatory regime flowing from the political agreement between the United States and China legally binding, and to extend that legal regime to other states. By broadening the regime to other countries, even those with no frontier model capability, but in which some firms might be tempted to run covert operations, the U.N. Security Council resolution would make both the restraint and the verification features of the model more effective—and thereby provide further mutual assurance and thus stability within the stag hunt dynamic. Finally, the resolution could establish or endorse the international standards and inspection body discussed above as a subsidiary organ of the U.N., as it did in the creation of UNSCOM with Resolution 687 (1991), the U.N. Special Commission established to verify and oversee Iraq’s elimination of prohibited weapons after the Gulf War.

U.N. Security Council resolution 1540 of 2004 serves as another precedent for this kind of move. Prompted by fears after 9/11 that non-state actors might acquire weapons of mass destruction, resolution 1540 imposed three sets of obligations on all members of the U.N.: states must refrain from supporting non-state actors that attempt to develop, acquire, possess, transport, transfer or use such weapons; they must adopt and enforce effective laws prohibiting non-state actors from doing any of these things; and they must establish domestic controls over related materials, including developing accounting and physical security systems, and border, trans-shipment, and export controls. It also created a committee and called upon member states to submit reports on compliance. 

The Security Council resolution contemplated here shares many of the same features, namely the imposition of an obligation to enact domestic law and regulation to control certain technology, and establishment of reporting regimens. It does go further in imposing the obligation to accept mandatory inspection and verification mechanisms. But if the United States and China both supported such a resolution, it is quite feasible that it could be passed. Russia is the most likely veto among the other three permanent members, but Russia does not yet have any frontier model development and so has no directly related motive to object. If it were passed, not only would it impose the foregoing obligations, but non-compliance and referral from the standards body could result in Article 41 measures under the Chapter VII resolution.

Conclusion

The executives pleading for restraint while still racing ahead, and the presidents refusing to provide any constraints, are making the same mistake: they are operating on the flawed premise that the AI race is a winner-takes-all contest, which is governed by the logic of the prisoner’s dilemma. Under that logic, cooperating on mutual restraint is not rational. But the experts warn that the reward for being first to develop AGI or RSI in insecure conditions may be an AI system that no one can control, with potentially catastrophic risks for everyone. Under that premise, the game-theory dynamic that best explains the situation is Rousseau’s stag hunt or assurance game—the logic of which dictates that restraint is each side’s best response to the other’s restraint, and cooperation is the stable equilibrium, provided each can verify that the other is cooperating. That is a problem of assurance, not of greed, and assurance can be produced by the right regulatory framework. 

Compute levels, weights, operating environments, and capabilities can each be made subject to observation, instrumented reporting, and independent audit. The layers of the legal framework set out above, from bilateral agreement to Security Council resolution, are designed to impose those restraints under a licensing regime and turn those observations into a system of mutual verification and assurance. As President Ronald Reagan often quipped during Cold War disarmament negotiations, “trust, but verify.” Cooperation between Washington and Beijing is the rational choice, and it can be sustained through verification. But it must start with a shared recognition of the risk and the will to act while the game still has two players.

Filed Under

, , , , , , , , ,
Send A Letter To The Editor

DON'T MISS A THING. Stay up to date with Just Security curated newsletters: