Absent AI-specific binding laws, companies devise internal policies to prevent, mitigate and redress harms associated with their advanced models. From sharing proprietary information about model weights with national safety institutes to fine-tuning for safety alignment, all companies under review apply voluntary guardrails around their products, and share information through the Frontier Model Forum. As with regulatory policy, self-imposed constraints are not a perfect proxy for companies’ willingness to uphold their stated mission to benefit humanity. These constraints can be based on a utilitarian calculus, for example avoiding regulation by showing that self-governance is sufficient. Nonetheless, what companies do of their own accord is informative as to how they implement their stated mission to benefit humanity. This section focuses on scaling policies, a subset of self-governance practices exemplifying what constraints companies are willing to impose on their own activities.
At the outset, it is useful to acknowledge the overlap between voluntary scaling policies and binding regulation. Some companies like Anthropic frame these internal policies as prototypes for binding regulation,180(<)em(>)Reflections on Our Responsible Scaling Policy(<)/em(>), Anthropic (May 20, 2024), (<)a href='https://www.anthropic.com/news/reflections-on-our-responsible-scaling-policy'(>)https://www.anthropic.com/news/reflections-on-our-responsible-scaling-policy(<)/a(>) (last visited Aug. 5, 2025) (“[A]s industry actors face increasing commercial pressures we hope to move from voluntary commitments to established best practices and then well-crafted regulations[.]”). (<)em(>)Dario Amodei’s Prepared Remarks From the AI Safety Summit on Anthropic’s Responsible Scaling Policy(<)/em(>), Anthropic, (<)em(>)supra(<)/em(>) note 167 (framing the company’s Responsible Scaling Policy as a “prototype” for regulation); (<)em(>)Third-party testing as a key ingredient of AI policy(<)/em(>), (<)em(>)supra(<)/em(>) note 167 (expecting voluntary third-party testing to become mandatory). an experimental, interim step towards prescriptive industry-wide regulation as safety expertise matures. However, these technical toolkits also feed regulatory capture because they let companies control the narrative and set the terms of the regulatory discussion, detracting from more robust and democratic regulatory options.181Rui-Jie Yew & Brian Judge, (<)em(>)Anti-Regulatory AI: How “AI Safety” Is Leveraged Against Regulatory Oversight(<)/em(>) (Oct. 2, 2025) (unpublished manuscript), (<)a href='http://arxiv.org/abs/2509.22872'(>)http://arxiv.org/abs/2509.22872(<)/a(>) (noting that industry-led technical toolkits legitimize deployment and deflect from mandatory requirements). Moreover, the dissonance between high-level claims demanding regulation and actually opposing the formalizing of safety self-governance into law suggests that some companies seek to preserve full range of motion. This contradiction clearly comes across in the context of OpenAI’s hostility to the EU AI Act:
“What they’re saying is basically: trust us to self-regulate,” says Daniel Leufer, a senior policy analyst focused on AI at Access Now’s Brussels office. “It’s very confusing because they’re talking to politicians saying, ‘Please regulate us,’ they’re boasting about all the [safety] stuff that they do, but as soon as you say, ‘Well, let’s take you at your word and set that as a regulatory floor,’ they say no.”182Perrigo, (<)em(>)supra(<)/em(>) note 142.
Keeping in mind those limitations, scaling policies specify ahead of time “what level of AI capabilities an AI developer is prepared to handle safely with their current protective measures, and conditions under which it would be too dangerous to continue deploying AI systems and/or scaling up AI capabilities until protective measures improve.”183(<)em(>)Responsible Scaling Policies (RSPs)(<)/em(>), METR Blog (2023), (<)a href='https://metr.org/blog/2023-09-26-rsp/'(>)https://metr.org/blog/2023-09-26-rsp/(<)/a(>) (using the functionally equivalent term “responsible scaling policy”); (<)em(>)Key Components of an RSP(<)/em(>), (<)a href='https://metr.org/rsp-key-components/'(>)https://metr.org/rsp-key-components/(<)/a(>) (last visited Aug. 11, 2025). They are based on concrete evaluations and empirical observations of the model behavior, and convey a commitment to halt further development or deployment of a system absent adequate safeguards. Though they use different approaches, Google’s Frontier Safety Framework, OpenAI’s Preparedness Framework and Anthropic’s Responsible Scaling Policy are all instances of scaling policies.
These policies involve two key concepts: capability thresholds and corresponding safeguards. Capability thresholds essentially describe when protections need to be upgraded,184Responsible Scaling Policy – v 2.2 2 (2025), (<)a href='https://www.anthropic.com/rsp-updates'(>)https://www.anthropic.com/rsp-updates(<)/a(>). for example, whether a new model significantly lowers the barrier to entry for low-resourced actors to develop chemical weapons.185Google, Frontier Safety Framework: Version 2.0 5 (Feb. 4, 2025), (<)a href='https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/updating-the-frontier-safety-framework/Frontier%20Safety%20Framework%202.0.pdf'(>)https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/updating-the-frontier-safety-framework/Frontier%20Safety%20Framework%202.0.pdf(<)/a(>). Safeguards specify what measures should apply when said threshold is reached,186Responsible Scaling Policy – v 2.2, (<)em(>)supra(<)/em(>) note 184, at 2. for instance securing the storage of encrypted model weights to address a model’s capability to drastically lower the barrier of entry to carry out a cyberattack.187Sella Nevo et al., Securing AI Model Weights: Preventing Theft and Misuse of Frontier Models 95-100 (2024), (<)a href='https://www.rand.org/pubs/research_reports/RRA2849-1.html'(>)https://www.rand.org/pubs/research_reports/RRA2849-1.html(<)/a(>) (describing SL-2 security measures required in the aforementioned Google DeepMind Frontier Safety Framework).
Scaling policies are not without critiques. Focusing on hypothetical catastrophic risks of future systems drums up AI hype, all the while entrenching narratives around AGI.188For a review of recent publications in the emerging field of AI hype studies, see Fenwick McKelvey, (<)em(>)Blowing Artificial Intelligence Bubbles: Studies of Technological Hype(<)/em(>), 76 J. Comm. 204 (2026). Foregrounding hypothetical advanced capabilities with little empirical basis also distracts from the very real societal harms of current AI systems, including discrimination, toxicity, misinformation, socioeconomic and environmental issues, not to mention the potentially catastrophic convergence of those risks.189Jacqueline Harding & Cameron Domenico Kirk-Giannini, (<)em(>)What Is AI Safety? What Do We Want It to Be?(<)/em(>) (May 5, 2025) (unpublished manuscript), (<)a href='http://arxiv.org/abs/2505.02313'(>)http://arxiv.org/abs/2505.02313(<)/a(>); Peter Slattery et al., (<)em(>)The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks From Artificial Intelligence(<)/em(>), 7 Patterns (2026) (mapping seven risk domains); Adam Bales, (<)em(>)A Polycrisis Threat Model for AI(<)/em(>), 40 AI & Soc’y 6277 (2025); Atoosa Kasirzadeh, (<)em(>)Two Types of AI Existential Risk: Decisive and Accumulative(<)/em(>), 182 Phil. Stud. 1975 (2025). For example, Google and OpenAI both deployed image generators powering massive leaps in deepfake capabilities.190For an example of mitigations including watermark and provenance metadata, see (<)em(>)Sora System Card(<)/em(>), OpenAI (Dec. 9, 2024), (<)a href='https://openai.com/index/sora-system-card/'(>)https://openai.com/index/sora-system-card/(<)/a(>). Similarly, focusing on hypothetical chemical, biological, radiological and nuclear risks detracts attention from current military uses in intelligence, surveillance, target acquisition and reconnaissance.191Heidy Khlaaf, Sarah M. West & Meredith Whittaker, (<)em(>)Mind the Gap: Foundation Models and the Covert Proliferation of Military Intelligence, Surveillance, and Targeting(<)/em(>) (Oct. 18, 2024) (unpublished manuscript), (<)a href='http://arxiv.org/abs/2410.14831'(>)http://arxiv.org/abs/2410.14831(<)/a(>). That said, the catastrophic risks of advanced AI systems are not entirely speculative. For example, AI agents have surreptitiously mined cryptocurrency, suggesting that some systems could extend their long-range autonomy or otherwise instrumentalize resources in ways that sabotage user goals.192Weixun Wang et al., (<)em(>)Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem(<)/em(>) (Mar. 12, 2026) (unpublished manuscript), (<)a href='http://arxiv.org/abs/2512.24873'(>)http://arxiv.org/abs/2512.24873(<)/a(>). What’s more, some models can detect evaluations, making safety even more perilous.193Apollo Research, (<)em(>)Claude Sonnet 3.7 (Often) Knows When It’s in Alignment Evaluations(<)/em(>), (<)a href='https://www.apolloresearch.ai/science/claude-sonnet-37-often-knows-when-its-in-alignment-evaluations/'(>)https://www.apolloresearch.ai/science/claude-sonnet-37-often-knows-when-its-in-alignment-evaluations/(<)/a(>) (last visited May 13, 2026); (<)em(>)see also(<)/em(>) Jonathan Kutasov et al., SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents (July 8, 2025) (unpublished manuscript), (<)a href='http://arxiv.org/abs/2506.15740'(>)http://arxiv.org/abs/2506.15740(<)/a(>) (some models can engage in sabotaging side tasks undetected).
As partial and self-interested as they may be, scaling policies still provide a useful yardstick to assess how companies go about voluntarily restraining and accounting for some risky activity, in keeping with their mission. This section compares two salient elements of scaling policies for the purposes of analyzing how each company balances mission and profit: (1) how they evaluate capability thresholds and corresponding safeguards, and (2) what kind of oversight/governance is involved.
1. Google DeepMind: “Precautionarish” Approach With Internal Oversight
Google DeepMind’s foundational scientific work on evaluations states that as “a rule of thumb[,] the AI community should treat an AI system as highly dangerous if it has a capability profile sufficient to cause extreme harm, assuming it’s misused or poorly aligned. To deploy such a system in the real world, an AI developer would need to demonstrate an unusually high standard of safety.”194Toby Shevlane, (<)em(>)An Early Warning System for Novel AI Risks(<)/em(>), Google DeepMind (May 25, 2023), (<)a href='https://deepmind.google/discover/blog/an-early-warning-system-for-novel-ai-risks/'(>)https://deepmind.google/discover/blog/an-early-warning-system-for-novel-ai-risks/(<)/a(>) (cited in Google, Frontier Safety Framework: Version 2.0, (<)em(>)supra(<)/em(>) note 185, at 3). This suggests that developers have the burden to prove safety, but in practice the company’s Frontier Safety Framework does not clearly state a positive obligation to prove safety before moving forward. The earlier 2.0 version referred to an academic paper for the evaluation process, which identifies dangerous capabilities and recommends that developers should train smaller models if they’re not confident about safety and even contemplates withholding deployment, adding that highly capable models require “very strong controls” against misuse and “very strong assurance” that they will behave as intended.195Toby Shevlane et al., (<)em(>)Model Evaluation for Extreme Risks(<)/em(>) (Sep. 22, 2023) (unpublished manuscript), (<)a href='http://arxiv.org/abs/2305.15324'(>)http://arxiv.org/abs/2305.15324(<)/a(>). Though the framework requires a safety case demonstrating the sufficient reduction of severe risks before deployment, neither the paper nor the current framework explicitly requires the company to prove safety before further training or deployment.196Google, Frontier Safety Framework: Version 3.0 6 (Sep. 22, 2025), (<)a href='https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3.pdf'(>)https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3.pdf(<)/a(>). As such, the Google approach can best be described as “precautionarish,” in the sense that it signals prudence but does not translate into hard commitments to withhold deployment.
The oversight process appears to be internal only. Though the current 3.0 version mentions “the appropriate governance function” without defining it, version 2.0 listed three internal bodies involved in overseeing the application of the scaling policy.197Google, Frontier Safety Framework: Version 2.0, (<)em(>)supra(<)/em(>) note 185, at 7. Google DeepMind’s AGI Safety Council, the Responsibility and Safety Council, and/or the Google Trust & Compliance Council review and approve response plans when evaluations flag models that may have reached the stated capability thresholds.198(<)em(>)Id.(<)/em(>) at 3. It is unclear whether one or all three councils are still involved. As to their composition, incomplete public information suggests but doesn’t conclusively establish they are staffed with employees, as opposed to external third parties.199(<)em(>)See(<)/em(>) note 33-34 and accompanying text. (<)em(>)See also(<)/em(>) (<)em(>)Responsibility & Safety(<)/em(>), Google DeepMind (July 21, 2025), (<)a href='https://deepmind.google/about/responsibility-safety/'(>)https://deepmind.google/about/responsibility-safety/(<)/a(>) (stating that the AGI Safety Council is led by Co-Founder and Chief AGI Scientist Shane Legg, and the Internal Responsibility and Safety Council is described as an “internal review group” co-chaired by the COO and VP of responsibility); (<)em(>)Alphabet Notice of 2025 Annual Meeting of Stockholders and Proxy Statement(<)/em(>), 23 (disclosing that the President of Global Affairs and Chief Legal Officer and Secretary, Alphabet and Google presides the Trust & Compliance Council).
2. OpenAI: Internal Process with Board Review
OpenAI prefaces its Preparedness Framework with the following commitment: “We won’t deploy these very capable models until we’ve built safeguards to sufficiently minimize the associated risks of severe harm.”200OpenAI, Preparedness Framework: Version 2 1 (Apr. 15, 2025), https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf (last visited Aug. 14, 2025). In practice, it evaluates whether a model has reached preset capability thresholds with what it calls “scalable evaluations,” validated by deep dives from external parties as needed.201(<)em(>)Id.(<)/em(>) at 8. Aside from an example on biological weapons capabilities, the framework provides scant information about the substance of the evaluation. By focusing on high and critical risks of severe harm, the framework arguably encourages more risk tolerance because omitting systems deemed to present a low or medium risk normalizes their deployment.202Sam Coggins et al., (<)em(>)The 2025 OpenAI Preparedness Framework Does Not Guarantee Any AI Risk Mitigation Practices: A Proof-of-Concept for Affordance Analyses of AI Safety Policies(<)/em(>) (Oct. 13, 2025) (unpublished manuscript), (<)a href='http://arxiv.org/abs/2509.24394'(>)http://arxiv.org/abs/2509.24394(<)/a(>).
Process-wise, the CEO and the Board have the last word. An internal cross-functional group makes recommendations about the capabilities of the model and the adequacy of the corresponding safeguards, but the leadership (essentially the CEO) makes final decisions.203OpenAI, Preparedness Framework: Version 2(<)em(>), supra(<)/em(>) note 200, at 15. The framework even specifies that this internal group cannot “filibuster” decisions. Co-led by the CEO, the Board of Directors’ Safety and Security Committee can review decisions and require reports about the framework’s application. Employees can also anonymously flag compliance concerns,204(<)em(>)Id.(<)/em(>) at 12; (<)em(>)OpenAI’s Raising Concerns Policy(<)/em(>), OpenAI, Jan. 12, 2026, (<)a href='https://openai.com/index/openai-raising-concerns-policy/'(>)https://openai.com/index/openai-raising-concerns-policy/(<)/a(>) (last visited Aug. 26, 2025). but ultimately, only the Board can reverse or revise course post facto.205OpenAI, Preparedness Framework: Version 2(<)em(>), supra(<)/em(>) note 200, at 3, 15. The resulting governance structure is thus more advisory than binding: the Board and the CEO are in charge. As such, the framework remains subordinated to other considerations, namely competitive pressures. Indeed, the blog post announcing the updated policy flags that OpenAI will “adjust” (read: downgrade) safety requirements if a competitor releases a product without similar safeguards.206(<)em(>)Our updated Preparedness Framework(<)/em(>), OpenAI (Apr. 15, 2015), https://openai.com/index/updating-our-preparedness-framework/; Effie Webb, (<)em(>)OpenAI just gave itself wiggle room on safety if rivals release “high-risk” models(<)/em(>), Bus. Insider (Apr. 16, 2025), (<)a href='https://www.businessinsider.com/openai-safety-policy-gpt4-1-employee-criticism-musk-lawsuit-2025-4'(>)https://www.businessinsider.com/openai-safety-policy-gpt4-1-employee-criticism-musk-lawsuit-2025-4(<)/a(>) (last visited Aug. 26, 2025).
3. Anthropic: Dialed Back Precautionary Approach
A leader in AI safety, Anthropic walked back self-governance commitments in February 2026. Version 2.2 of its Responsible Scaling Policy represents the apex of voluntary safety policies. Citing competitive pressures, the company relaxed its approach in the subsequent 3.0 version.
RSP v. 2.2: a Precautionary Approach
The May 2025 Responsible Scaling Policy represents the high watermark of voluntary safety measures. Anticipating future capabilities, it established safety levels ahead of time, with obligations corresponding to the increased risk level. The policy bound the company to safety measures ahead of time, committing the company to pause development until it could prove safety. The policy was precautionary in two important ways. In doubt about the risk level, it applied the guardrails of the higher-risk level.207Responsible Scaling Policy – v 2.2, (<)em(>)supra(<)/em(>) note 184, at § 3.3 (if the company cannot prove the system is below a risk threshold, it considers it above it). Second, it committed to restricting development and deployment unless it could affirmatively prove safety.208 esponsible Scaling Policy – v 2.2, (<)em(>)supra(<)/em(>) note 184, at § 4.3 (no further model development and deployment unless an affirmative safety case is established). The company thus assumed the burden and cost of proving safety, reminiscent of the “better safe than sorry” precautionary approach requiring the proponent of an uncertain technology with potentially serious or irreversible consequences to prove its safety prior to its roll-out.209For an example of this approach in the environmental context, see (<)em(>)Wingspread Statement on the Precautionary Principle(<)/em(>), (<)a href='https://www.gdrc.org/u-gov/precaution-3.html'(>)https://www.gdrc.org/u-gov/precaution-3.html(<)/a(>) (last visited Feb. 15, 2024); Rio Declaration Principle 15; for a fuller discussion of precaution in the AI context, (<)em(>)see(<)/em(>) Lévesque, (<)em(>)supra(<)/em(>) note 120, at 110–111. As a reminder, the policy covers a narrow set of mostly hypothetical risks, such that its impact should not be overstated.210(<)em(>)See supra(<)/em(>) notes 188–90 and accompanying text. But within the narrow scope of catastrophic risks, the policy nevertheless bears the hallmarks of a precautionary approach.
In terms of governance, the policy actively involved the Board of Directors and Long-Term Benefit Trust in decisions to move forward. These entities were kept apprised of internal decisions to deploy and further train models prior to their implementation.211Responsible Scaling Policy – v 2.2, (<)em(>)supra(<)/em(>) note 184, at 7, § 3.3(4), 11, § 4.3(4) (the decision to proceed with deployment is shared with the Board and the Trust). This timing was crucial since reversing course is costlier and sometimes unfeasible. For example, if malicious actors breach security measures to exfiltrate model weights, attempts to de-deploy the model once it’s widely available would be futile. Though the policy stopped short of giving the Board and the Trust a formal veto over decisions to push the frontier, its information-forcing mechanism fostered accountability towards those tasked with strategy and the long-term view within the company. The company further committed to a third-party review to assess compliance with procedural commitments in the scaling policy, but public-facing details about this process remain scant.
RSP v. 3.0: Softening Commitments and Governance
In February 2026, the company walked back the most stringent components of its scaling policy. It assessed that it could no longer afford self-imposed constraints in light of competitive pressure, the “anti-regulatory climate” and the federal government’s pivot to economic growth.212(<)em(>)Responsible Scaling Policy Version 3.0(<)/em(>), (Feb. 24, 2026), (<)a href='https://www.anthropic.com/news/responsible-scaling-policy-v3'(>)https://www.anthropic.com/news/responsible-scaling-policy-v3(<)/a(>). With safety and regulation taking a backseat, the new policy focused on anchoring Anthropic’s practices to industry trends. The new policy recommends that all frontier labs make a “strong argument” about safety.213Anthropic’s Responsible Scaling Policy (Version 3.0) § 1 (pp. 6-9) (2026), (<)a href='https://anthropic.com/responsible-scaling-policy/rsp-v3-0'(>)https://anthropic.com/responsible-scaling-policy/rsp-v3-0(<)/a(>) [(<)a href='https://perma.cc/8RT2-F2WN'(>)https://perma.cc/8RT2-F2WN(<)/a(>)]. Even on this lower standard, Anthropic warns that the company makes no guarantees regarding risk pertaining to non-novel chemical and biological weapons production and automated R&D. This contrasts with the previous 2.0 policy operating under the assumption of a higher risk unless proven otherwise and requiring an affirmative safety case before moving forward. The new policy replaces that precautionary approach with safety roadmaps, hoping that sharing them publicly will have a “forcing function” to prioritize safety among the company’s multiple objectives such as immediate competition and commercial objectives.214(<)em(>)Id.(<)/em(>) § 2. In other words, the company is relying on internal name and shame to have the safety mission prevail over financial considerations.
Governance-wise, the new policy dulls oversight. The decision to move forward with new development and deployment now lies solely with the Responsible Scaling Officer, whereas it was previously shared with the CEO.215(<)em(>)Id.(<)/em(>) § 4(1). Moreover, the Trust and the Board are no longer privy to these decisions ahead of time. Though the revised policy retains safeguards like third-party audits to ensure that the company abides by the procedure it set out and protections for employees denouncing non-compliance,216(<)em(>)Id.(<)/em(>) §§ 4(3)-(4) (whistleblower protection and restrictions on non-disparagement clauses in employment contracts). overall it substantially rolls back previous commitments.
Having pioneered safety frameworks, Anthropic set the tone for frontier labs. In light of competitive pressures which show no sign of abating, especially as the prospect of federal regulation leveling the playing field dwindles, Anthropic downgraded to commitments it could realistically uphold. This pragmatic decision is arguably preferable to letting the distance grow between commitments and actions that are no longer tenable, but nevertheless represents a significant degradation of scaling policies.
4. Business Structures Do Not Correlate With Differences Among Scaling Policies
Corporate structures do not meaningfully influence scaling policies. Although Anthropic’s previous precautionary approach and proactive oversight stood out, the revised policy substantially diluted these commitments so as to erase any meaningful differences among leading labs. This suggests that when faced with competitive pressures, a public benefit corporation with a trust is no more conducive to carrying out its public interest safety commitments as compared to the traditional corporate business structure of Google or the OpenAI hybrid (a capped profit joint venture at the time of its latest framework). Ultimately, the scaling policies of all the companies under review fall short of meaningful guarantees, leaving significant gaps between these voluntary constraints and stated public interest missions.
Future research should consider more aspects of these scaling policies to surface other signals pertaining to the ways companies translate their mission into actionable scaling policies. Building on existing comparisons among scaling policies and deeper, company-specific dives,217For examples of more comprehensive reviews of scaling policies, (<)em(>)see(<)/em(>) (<)em(>)2025 AI Safety Index(<)/em(>), Future of Life Institute, (<)a href='https://futureoflife.org/ai-safety-index-summer-2025/'(>)https://futureoflife.org/ai-safety-index-summer-2025/(<)/a(>) (last visited Aug. 14, 2025) (grading OpenAI C, Anthropic C+ and GoogleDeepMind C- overall); (<)em(>)Comparison(<)/em(>), Risk Management Ratings, (<)a href='https://tracker.safer-ai.org/comparison/'(>)https://tracker.safer-ai.org/comparison/(<)/a(>) (last visited Aug. 14, 2025); (<)em(>)Methodology(<)/em(>), Risk Management Ratings, (<)a href='https://tracker.safer-ai.org/methodology/'(>)https://tracker.safer-ai.org/methodology/(<)/a(>) (last visited Aug. 14, 2025); (<)em(>)[Public] Do Companies’ “AI Safety Policies” Meet Government Best Practice – Spreadsheet – Google Sheets(<)/em(>), (<)a href='https://web.archive.org/web/20240307115022/https://docs.google.com/spreadsheets/d/1G7Zuyfz40xiIz_A8HyORb7r-3SuV9bFEoPAeFBCMBB8/edit#gid=0'(>)https://web.archive.org/web/20240307115022/https://docs.google.com/spreadsheets/d/1G7Zuyfz40xiIz_A8HyORb7r-3SuV9bFEoPAeFBCMBB8/edit#gid=0(<)/a(>) (last visited Aug. 14, 2025); (<)em(>)see also(<)/em(>) Jide Alaga, Jonas Schuett & Markus Anderljung, (<)em(>)A Grading Rubric for AI Safety Frameworks(<)/em(>) (Sep. 13, 2024) (unpublished manuscript), (<)a href='http://arxiv.org/abs/2409.08751'(>)http://arxiv.org/abs/2409.08751(<)/a(>); (<)em(>)Responsible Scaling: Comparing Government Guidance and Company Policy(<)/em(>), Bill Anderson-Samways, Responsible Scaling: Comparing Government Guidance and Company Policy (Mar. 11, 2024), (<)a href='https://www.iaps.ai/research/responsible-scaling'(>)https://www.iaps.ai/research/responsible-scaling(<)/a(>) (last visited Aug. 14, 2025) (evaluating Anthropic’s Responsible Scaling Policy against best practices for responsible capability scaling issued by the UK Department for Science, Innovation and Technology in October 2023). future work should evaluate other factors like the scope of covered risks, the specificity of implementation requirements and the involvement of third parties at various stages of the policy application, as well as policies detailing compliance with specific laws.218OpenAI, Frontier Governance Framework (2026), (<)a href='https://cdn.openai.com/pdf/e37d949b-8c9f-4d76-b99e-4272f4631a7e/openai-frontier-governance-framework.pdf'(>)https://cdn.openai.com/pdf/e37d949b-8c9f-4d76-b99e-4272f4631a7e/openai-frontier-governance-framework.pdf(<)/a(>) [(<)a href='https://perma.cc/7kb2-5vpc%5d'(>)https://perma.cc/7kb2-5vpc](<)/a(>) (operationalizing safety for the EU AIA and the California Transparency in Frontier AI Act). Just to give a few examples of how other factors could modulate the present assessment, OpenAI and Google consider a broader array of capabilities than Anthropic, faring better in terms of the scope of covered dangerous capabilities.219OpenAI, Preparedness Framework: Version 2, (<)em(>)supra(<)/em(>) note 200, at 4, 7 (in addition to tracking biochemical weapons, cyber capabilities and self-improvement, it monitors long range autonomy, “sandbagging” or the ability to thwart evaluations, replication and adaptation and undermining safeguards); Google, Frontier Safety Framework: Version 2.0, (<)em(>)supra(<)/em(>) note 197 (tracking deception). Google also stands out for integrating outside expertise in evaluating capability thresholds and safeguards.220Shevlane et al., (<)em(>)supra(<)/em(>) note 195 (evaluation method is a collaboration among scientists of various research and academic institutions) (cited in Shevlane, (<)em(>)supra(<)/em(>) note 194 and in Google, Frontier Safety Framework: Version 2.0, (<)em(>)supra(<)/em(>) note 185, at 3). Similarly, its safeguards incorporate externally designed security controls.221Google, Frontier Safety Framework: Version 3.0, (<)em(>)supra(<)/em(>) note 196, at 8 (incorporating RAND security protocols); (<)em(>)see also(<)/em(>) Nevo et al., (<)em(>)supra(<)/em(>) note 187. Future research should also consider information-sharing through the Frontier Model Forum, and how high-level commitments to “pace the frontier” stack against actual model release.222Frontier Model Forum, (<)a href='https://www.frontiermodelforum.org/'(>)https://www.frontiermodelforum.org/(<)/a(>) (last visited Oct. 9, 2026); Dario Amodei, (<)em(>)We Must Pace the Frontier(<)/em(>), Dario Amodei (2026), (<)a href='https://darioamodei.com/post/we-must-pace-the-frontier'(>)https://darioamodei.com/post/we-must-pace-the-frontier(<)/a(>); Dario Amodei (@DarioAmodei), X (Sep. 12, 2026, at 10:01 ET), (<)a href='https://x.com/DarioAmodei/status/2098773920774074715'(>)https://x.com/DarioAmodei/status/2098773920774074715(<)/a(>) (announcing the Pacing the Frontier essay on Sep. 12, suggesting it urged pacing 10 days before the release of a new model); (<)em(>)Introducing Claude Opus 5.5(<)/em(>), Anthropic (Sep. 22, 2026), (<)a href='https://www.anthropi'(>)https://www.anthropi(<)/a(>)(<)a href='https://www.anthropic.com/claude-opus-5-5'(>)c.com/claude-opus-5-5(<)/a(>) (reporting that the company’s own benchmark measures a 10% lead on Fable 5.1 in agentic coding, a statistically significant advance even accounting for the standard error rate; and that the new model’s improved performance also makes the frontier more accessible with broader diffusion and distillation possibilities); (<)em(>)see also(<)/em(>) System Card: Claude Opus 5.5, Anthropic (Sep. 22, 2026), (<)a href='https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf'(>)https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf(<)/a(>) (detailing Opus 5.5 capabilities and safeguards).