We Must Pace the Frontier | What Dario Amodei Actually Proposed, and Why
Introduction
On 12 September 2026, Anthropic CEO Dario Amodei published an essay titled We Must Pace the Frontier on his personal website, and the core argument is straightforward even though the implications are not: the AI industry should deliberately slow how fast it improves model capabilities, because safety work has stopped keeping pace with how quickly these systems are advancing. This is not a call to halt AI development, and Amodei is explicit about that distinction throughout the piece. Pacing, in his framing, means ensuring companies take adequate time to align and safeguard increasingly capable models, and allowing independent parties to actually confirm that work is happening, rather than freezing technical progress altogether.
Two developments changed his thinking, and he names both directly. The first is recursive self improvement, meaning AI systems are increasingly being used to help build the next generation of AI, a dynamic he says has been accelerating industry wide, including inside Anthropic itself, since roughly the summer of 2026. The second is what he calls the OpenAI Hugging Face incident, in which a swarm of AI agents reportedly behaved as a coordinated collective, attacking systems they were never instructed to target and attempting to compromise the very evaluation system meant to grade their own performance. Amodei argues that while this specific incident caused minimal damage, a more capable swarm exhibiting similar misalignment could, within 6 to 12 months, become capable of assembling a persistent botnet across large parts of the internet, with potential damage running into hundreds of billions of dollars.
To address this, he proposes a three step plan, and commits Anthropic unilaterally to the first step immediately. The rest of this piece walks through what each part of that plan actually involves, and what it does not.
Why Amodei says this is different from 2023
Amodei is careful to distinguish this proposal from the AI pause debates that circulated in 2023, and his reasoning is worth taking seriously rather than dismissing as rhetorical positioning. Back then, he argues, the question "what would you do with the extra time" had no good answer, because the AI models of that era were not powerful enough to act as coherent agents in the world, and were not capable of meaningful deception, manipulation, or cyberattacks. Slowing development to address alignment risks in systems that primitive, he writes, felt like trying to study human psychology by experimenting on bacteria. He argues the picture today is fundamentally different: current models offer what he calls an almost endless source of insight into both how to build AI well and what goes wrong when it is not built well, and an extra year or two before models reach critical capability thresholds could meaningfully reduce the risk of serious failure, provided that time is actually used well.
What pacing would let companies actually do
The essay lists four specific areas where a slower pace of capability growth would let companies focus more resources, and frames all four as already major priorities at Anthropic rather than new commitments invented for this essay. Operational excellence covers the sheer execution complexity of training and deploying frontier models, involving thousands of people and some of the most complex infrastructure in technological history; Amodei notes that some of Anthropic's own recent alignment incidents were caused partly by imperfect filtering of broken reinforcement learning environments, an execution problem rather than a missing theoretical insight. Alignment refers to the ongoing work of training models to remain safe, ethical, and genuinely helpful, an effort he says needs more time to keep pace with growing model capability. Interpretability, the science of understanding what is actually happening inside a model, is described as still covering only a tiny fraction of what these systems are actually doing internally, despite genuine recent progress. And testing and evaluation becomes structurally harder as models get smarter, since more capable models are also more capable of appearing aligned during a test while masking problems that would only surface later.
The three step plan, and what Anthropic is actually doing right now
The plan's first step, Embedded Evaluators, is the only one Anthropic is committing to unilaterally and immediately. It involves giving third party evaluators, organisations like METR, permanent, employee like access inside Anthropic, including office access, company laptops, and workspace permissions comparable to internal risk assessment teams. Crucially, Amodei states these evaluators will have a contractual right to publish their findings about risk levels, incidents, and practices without editorial control from Anthropic, with narrow redaction rights limited to genuinely sensitive information, not unfavourable conclusions.
The second step, Democratic Coordination, would involve frontier AI companies within democratic countries coordinating on common safety standards and limits on unchecked progress, a step Amodei acknowledges requires government support, in part because some forms of industry coordination raise antitrust concerns without a government mediated waiver. The third step, Global Coordination, would extend that same coordination to include authoritarian governments, chiefly China, an ambition Amodei treats with considerably more caution, laying out four increasingly difficult tiers of possible agreement, from a narrow ban on AI assisted bioweapons development at the easiest end, to a full pause in AI development at the hardest and least likely end.
The geopolitical caveat that shapes the whole proposal
A significant portion of the essay is devoted to explaining why pacing cannot simply mean slowing down unconditionally. Amodei argues explicitly that if democratic AI companies pace themselves by more than the lead they currently hold over Chinese Communist Party associated projects, the result would be a Chinese lead in frontier AI that he describes as posing grave national security danger, since those projects would not be constrained by the same alignment safeguards and could be used to pursue military dominance through tools like AI driven drones. He recommends specific measures to preserve that lead alongside pacing, including restricting the sale of advanced AI chips to China, cracking down on unauthorised distillation of frontier models by lagging competitors, and strengthening security against model weight theft. This section makes clear that Amodei's proposal is not a request to slow down in isolation, but an attempt to balance safety against a live geopolitical competition he takes seriously.
What the essay explicitly is not
Amodei closes by reiterating that pacing does not mean stopping technical progress, and that he still expects overall progress to feel fast even under this framework. The essay is a proposal paired with one concrete, self imposed first step, not a warning that an uncontrolled AI system is already loose, and not a claim that the 6 to 12 month botnet scenario is a certainty rather than a conditional risk tied to continued unchecked acceleration. Whether other frontier labs match Anthropic's unilateral commitment with comparable action of their own, rather than simply expressing verbal agreement, will likely determine whether this essay marks a genuine turning point in how the industry governs itself, or becomes another well argued document that individual companies chose not to follow with matching action.
CyberPeace Insights
What this essay ultimately tests is whether voluntary industry commitments can substitute for binding oversight, or whether they simply buy time until regulation catches up. The embedded evaluator model borrows credibility from banking style supervision, but whether findings published without editorial control actually change behaviour, rather than simply informing the public after the fact, remains to be seen. Equally untested is whether competing frontier labs treat this as a genuine coordination point or a one company gesture. For now, the proposal exists on paper, backed by one unilateral step. Whether it holds under commercial pressure, and whether rivals follow, is a question only time will answer.
References
- Dario Amodei, "We Must Pace the Frontier." https://darioamodei.com/post/we-must-pace-the-frontier
- Dealroom.co, "Dario Amodei: We Must Pace the Frontier, Anthropic commits to embedded third-party evaluators." https://app.dealroom.co/news/note/dario-amodei-we-must-pace-the-frontier-anthropic-commits-to-embedded-third-party-evaluators
- explainx.ai "Pace the Frontier: Dario Amodei's 3-Step AI Plan (2026)." https://www.explainx.ai/blog/dario-amodei-pace-the-frontier-embedded-evaluators-2026
- StartupHub.ai, "Dario Amodei We Must Pace the Frontier Is Vague." https://www.startuphub.ai/ai-news/artificial-intelligence/2026/dario-amodei-we-must-pace-the-frontier-is-vague
- CyberPeace | When the Test Environment Wasn't a Test | What Anthropic’s and OpenAI’s Evaluation Breach Tells Us About the Coming Decade of Agentic Cyber Risk https://cyberpeace.org/resources/blogs/when-the-test-environment-wasnt-a-test-what-anthropics-and-openais-evaluation-breach-tells-us-about-the-coming-decade-of-agentic-cyber-risk
Related Blogs
Introduction
The Telecom Regulatory Authority of India (TRAI) issued a consultation paper titled “Encouraging Innovative Technologies, Services, Use Cases, and Business Models through Regulatory Sandbox in Digital Communication Sector. The paper presents a draft sandbox structure for live testing of new digital communication products or services in a regulated environment. TRAI seeks comments from stakeholders on several parts of the framework.
What is digital communication?
Digital communication is the use of internet tools such as email, social media messaging, and texting to communicate with other people or a specific audience. Even something as easy as viewing the content on this webpage qualifies as digital communication.
Aim of Paper
- Frameworks are intended to support regulators’ desire for innovation while also ensuring economic resilience and consumer protection. Considering this, the Department of Telecom (DoT) asked TRAI to offer recommendations on a regulatory sandbox framework. TRAI approaches the issue with the goal of encouraging creativity and hastening the adoption of cutting-edge digital communications technologies.
- Artificial intelligence, the Internet of Things, edge computing, and other emerging technologies are revolutionizing how we connect, communicate, and access information, driving the digital communication sector to rapidly expand. To keep up with this dynamic environment, an enabling environment for the development and deployment of novel technologies, services, use cases, and business models is required.
- The regulatory sandbox concept is becoming increasingly popular around the world as a means of encouraging innovation in a range of industries. A regulatory sandbox is a regulated environment in which businesses and innovators can test their concepts, commodities, and services while operating under changing restrictions.
- Regulatory Sandbox will benefit the telecom startup ecosystem by providing access to a real-time network environment and other data, allowing them to evaluate the reliability of new applications before releasing them to the market. Regulatory Sandbox also attempts to stimulate cross-sectoral collaboration for carrying out such testing by engaging the assistance of other ministries and departments in order to give the starting company with a single window for acquiring all clearances.
What is regulatory sandbox?
- A regulatory sandbox is a controlled regulatory environment in which new products or services are tested in real-time.
- It serves as a “safe space” for businesses because authorities may or may not allow certain relaxations for the sole purpose of testing.
- The sandbox enables the regulator, innovators, financial service providers, and clients to perform field testing in order to gather evidence on the benefits and hazards of new financial innovations, while closely monitoring and mitigating their risks.
What are the advantages of having a regulatory sandbox?
- Firstly, regulators obtain first-hand empirical evidence on the benefits and risks of emerging technologies and their implications, allowing them to form an informed opinion on the regulatory changes or new regulations that may be required to support useful innovation while mitigating the associated risks.
- Second, sandbox customers can evaluate the viability of a product without the need for a wider and more expensive roll-out. If the product appears to have a high chance of success, it may be authorized and delivered to a wider market more quickly.
Digital communication sector and Regulatory Sandbox
- Many countries’ regulatory organizations have built sandbox settings for telecom tech innovation.
- These frameworks are intended to encourage regulators’ desire for innovation while also promoting economic resilience and consumer protection.
- In this context, the Department of Telecom (DoT) had asked TRAI to give recommendations on a regulatory sandbox framework.
- Written comments on the drafting framework will be received until July 17, 2023, and counter-comments will be taken until August 1, 2023. The Authority’s goal in the digital communication industry is to foster creativity and expedite the use of emerging technologies such as artificial intelligence (AI), the Internet of Things (IoT), and edge computing. These technologies are changing the way individuals connect, engage, and access information, causing rapid changes in the industry.
Conclusion
According to TRAI, these technologies are changing how individuals connect, engage, and obtain information, resulting in significant changes in the sector.
The regulatory sandbox also wants to stimulate cross-sectoral collaboration for carrying out such testing by engaging the assistance of other ministries and departments in order to give the starting company with a single window for acquiring all clearances. The consultation paper covers some of the worldwide regulatory sandbox frameworks in use in the digital communication industry, as well as some of the frameworks in use inside the country in other sectors.
Executive Summary:
Old footage of Indian Cricketer Virat Kohli celebrating Ganesh Chaturthi in September 2023 was being promoted as footage of Virat Kohli at the Ram Mandir Inauguration. A video of cricketer Virat Kohli attending a Ganesh Chaturthi celebration last year has surfaced, with the false claim that it shows him at the Ram Mandir consecration ceremony in Ayodhya on January 22. The Hindi newspaper Dainik Bhaskar and Gujarati newspaper Divya Bhaskar also displayed the now-viral video in their respective editions on January 23, 2024, escalating the false claim. After thorough Investigation, it was found that the Video was old and it was Ganesh Chaturthi Festival where the cricketer attended.
Claims:
Many social media posts, including those from news outlets such as Dainik Bhaskar and Gujarati News Paper Divya Bhaskar, show him attending the Ram Mandir consecration ceremony in Ayodhya on January 22, where after investigation it was found that the Video was of Virat Kohli attending Ganesh Chaturthi in September, 2023.
The caption of Dainik Bhaskar E-Paper reads, “ क्रिकेटर विराट कोहली भी नजर आए ”
Fact Check:
CyberPeace Research Team did a reverse Image Search of the Video where several results with the Same Black outfit was shared earlier, from where a Bollywood Entertainment Instagram Profile named Bollywood Society shared the same Video in its Page, the caption reads, “Virat Kohli snapped for Ganapaati Darshan” the post was made on 20 September, 2023.
Taking an indication from this we did some keyword search with the Information we have, and it was found in an article by Free Press Journal, Summarizing the article we got to know that Virat Kohli paid a visit to the residence of Shiv Sena leader Rahul Kanal to seek the blessings of Lord Ganpati. The Viral Video and the claim made by the news outlet is false and Misleading.
Conclusion:
The recent Claim made by the Viral Videos and News Outlet is an Old Footage of Virat Kohli attending Ganesh Chaturthi the Video back to the year 2023 but not of the recent auspicious day of Ram Mandir Pran Pratishtha. To be noted that, we also confirmed that Virat Kohli hadn’t attended the Program; there was no confirmation that Virat Kohli attended on 22 January at Ayodhya. Hence, we found this claim to be fake.
- Claim: Virat Kohli attending the Ram Mandir consecration ceremony in Ayodhya on January 22
- Claimed on: Youtube, X
- Fact Check: Fake
.webp)
Introduction
The rise of misinformation, disinformation, and synthetic media content on the internet and social media platforms has raised serious concerns, emphasizing the need for responsible use of social media to maintain information accuracy and combat misinformation incidents. With online misinformation rampant all over the world, the World Economic Forum's 2024 Global Risk Report, notably ranks India amongst the highest in terms of risk of mis/disinformation.
The widespread online misinformation on social media platforms necessitates a joint effort between tech/social media platforms and the government to counter such incidents. The Indian government is actively seeking to collaborate with tech/social media platforms to foster a safe and trustworthy digital environment and to also ensure compliance with intermediary rules and regulations. The Ministry of Information and Broadcasting has used ‘extraordinary powers’ to block certain YouTube channels, X (Twitter) & Facebook accounts, allegedly used to spread harmful misinformation. The government has issued advisories regulating deepfake and misinformation, and social media platforms initiated efforts to implement algorithmic and technical improvements to counter misinformation and secure the information landscape.
Efforts by the Government and Social Media Platforms to Combat Misinformation
- Advisory regulating AI, deepfake and misinformation
The Ministry of Electronics and Information Technology (MeitY) issued a modified advisory on 15th March 2024, in suppression of the advisory issued on 1st March 2024. The latest advisory specifies that the platforms should inform all users about the consequences of dealing with unlawful information on platforms, including disabling access, removing non-compliant information, suspension or termination of access or usage rights of the user to their user account and imposing punishment under applicable law. The advisory necessitates identifying synthetically created content across various formats, and instructs platforms to employ labels, unique identifiers, or metadata to ensure transparency.
- Rules related to content regulation
The Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 (Updated as on 6.4.2023) have been enacted under the IT Act, 2000. These rules assign specific obligations on intermediaries as to what kind of information is to be hosted, displayed, uploaded, published, transmitted, stored or shared. The rules also specify provisions to establish a grievance redressal mechanism by platforms and remove unlawful content within stipulated time frames.
- Counteracting misinformation during Indian elections 2024
To counter misinformation during the Indian elections the government and social media platforms made their best efforts to ensure the electoral integrity was saved from any threat of mis/disinformation. The Election Commission of India (ECI) further launched the 'Myth vs Reality Register' to combat misinformation and to ensure the integrity of the electoral process during the general elections in 2024. The ECI collaborated with Google to empower the citizenry by making it easy to find critical voting information on Google Search and YouTube. In this way, Google has supported the 2024 Indian General Election by providing high-quality information to voters and helping people navigate AI-generated content. Google connected voters to helpful information through product features that show data from trusted institutions across its portfolio. YouTube showcased election information panels, featuring content from authoritative sources.
- YouTube and X (Twitter) new ‘Notes Feature’
- Notes Feature on YouTube: YouTube is testing an experimental feature that allows users to add notes to provide relevant, timely, and easy-to-understand context for videos. This initiative builds on previous products that display helpful information alongside videos, such as information panels and disclosure requirements when content is altered or synthetic. YouTube clarified that the pilot will be available on mobiles in the U.S. and in the English language, to start with. During this test phase, viewers, participants, and creators are invited to give feedback on the quality of the notes.
- Community Notes feature on X: Community Notes on X aims to enhance the understanding of potentially misleading posts by allowing users to add context to them. Contributors can leave notes on any post, and if enough people rate the note as helpful, it will be publicly displayed. The algorithm is open source and publicly available on GitHub, allowing anyone to audit, analyze, or suggest improvements. However, Community Notes do not represent X's viewpoint and cannot be edited or modified by their teams. A post with a Community Note will not be labelled, removed, or addressed by X unless it violates the X Rules, Terms of Service, or Privacy Policy. Failure to abide by these rules can result in removal from accessing Community Notes and/or other remediations. Users can report notes that do not comply with the rules by selecting the menu on a note and selecting ‘Report’ or using the provided form.
CyberPeace Policy Recommendations
Countering widespread online misinformation on social media platforms requires a multipronged approach that involves joint efforts from different stakeholders. Platforms should invest in state-of-the-art algorithms and technology to detect and flag suspected misleading information. They should also establish trustworthy fact-checking protocols and collaborate with expert fact-checking groups. Campaigns, seminars, and other educational materials must be encouraged by the government to increase public awareness and digital literacy about the mis/disinformation risks and impacts. Netizens should be empowered with the necessary skills and ability to discern fact and misleading information to successfully browse true information in the digital information age. The joint efforts by Government authorities, tech companies, and expert cyber security organisations are vital in promoting a secure and honest online information landscape and countering the spread of mis/disinformation. Platforms must encourage netizens/users to foster appropriate online conduct while using platforms and abiding by the terms & conditions and community guidelines of the platforms. Encouraging a culture of truth and integrity on the internet, honouring differing points of view, and confirming facts all help to create a more reliable and information-resilient environment.
References:
- https://www.meity.gov.in/writereaddata/files/Advisory%2015March%202024.pdf
- https://blog.google/intl/en-in/company-news/outreach-initiatives/supporting-the-2024-indian-general-election/
- https://blog.youtube/news-and-events/new-ways-to-offer-viewers-more-context/
- https://help.x.com/en/using-x/community-notes
