MAIM (Mutually Assured AI Malfunction)

Overview

Mutually Assured AI Malfunction (MAIM (Mutually Assured AI Malfunction)ApproachMAIM (Mutually Assured AI Malfunction)Comprehensive reference on the MAIM deterrence framework proposed by Hendrycks, Schmidt, and Wang, covering the core theory, escalation ladder, major critiques from MIRI/RAND/IAPS, and developments...Quality: 55/100) is a proposed deterrence framework for managing great-power competition over advanced AI. Introduced by Dan HendrycksPersonDan HendrycksComprehensive reference biography of Dan Hendrycks (CAIS director), covering his academic career (GELU, MMLU, OOD detection), CAIS founding and funding (including $6.5M FTX, Open Philanthropy/Coeff...Quality: 19/100 (director of Center for AI SafetyOrganizationCenter for AI SafetyCAIS is a nonprofit research organization founded by Dan Hendrycks that has distributed compute grants to researchers, published technical AI safety papers including the representation engineering ...Quality: 42/100), Eric Schmidt (former Google CEO), and Alexandr Wang (Scale AI CEO) in their March 2025 paper Superintelligence Strategy (revised April 2025), MAIM proposes that any state's aggressive bid for unilateral AI dominance will be met with preventive sabotage by rivals.¹

The framework draws an analogy to nuclear Mutually Assured Destruction (MAD), but operates through preemption rather than retaliation. The authors argue that because — in their assessment — destabilizing AI projects are relatively easy to sabotage through interventions ranging from covert cyberattacks to kinetic strikes on datacenters, MAIM already describes the strategic picture AI superpowers find themselves in. The resulting stalemate could, in their view, postpone the emergence of SuperintelligenceConceptSuperintelligenceAI systems with cognitive abilities vastly exceeding human intelligenceQuality: 92/100, curtail loss-of-control scenarios, and undercut efforts to secure a strategic monopoly. These claims are contested by analysts who question whether sabotage would be technically effective, whether AI development is sufficiently observable, and whether the nuclear analogy holds in the AI context.

MAIM is one pillar of a broader three-part framework alongside nonproliferation (tracking AI chips and preventing rogue access) and competitiveness (guaranteeing domestic chip manufacturing capacity). The paper has generated extensive debate, with critiques from Machine Intelligence Research InstituteOrganizationMachine Intelligence Research InstituteComprehensive organizational history documenting MIRI's trajectory from pioneering AI safety research (2000-2020) to policy advocacy after acknowledging research failure, with detailed financial da...Quality: 50/100, RAND, and the Institute for AI Policy and StrategyOrganizationInstitute for AI Policy and StrategyAI policy think tank founded in 2024, headquartered in the Washington, DC metro area. Led by Jennifer Marron (Executive Director), Brianna Rosen (Research Director, Frontier Security), Amanda El-Da... raising concerns about observability, escalation risks, and the limits of the nuclear analogy.

Quick Assessment

MAIM's viability depends on a set of empirical conditions that analysts continue to contest. As of early 2026, no state has formally adopted MAIM as doctrine, and the framework has received mixed assessments from security analysts, deterrence theorists, and AI policy researchers.

An IAPS analysis (Delaney, December 2025) assigned approximately 70% probability that China would anticipate near-total loss of great-power status if the US achieves superintelligence unilaterally, ~60% probability that China would pursue escalating sabotage actions in response, and ~60% probability the US would negotiate a mutual slowdown rather than risk nuclear confrontation — yielding approximately 25% combined probability that all descriptive MAIM conditions are met.²

Risks Addressed

MAIM is proposed as a response to several interconnected risks in the development of advanced AI:

• AI Development Racing DynamicsRiskAI Development Racing DynamicsRacing dynamics analysis shows competitive pressure has shortened safety evaluation timelines by 40-60% since ChatGPT's launch, with commercial labs reducing safety work from 12 weeks to 4-6 weeks....Quality: 72/100: Competitive pressure between AI-developing states may incentivize shortcuts on safety and accelerate the timeline toward dangerous capabilities. MAIM's deterrence equilibrium is intended to slow the most destabilizing development efforts.
• Unilateral AI dominance: The authors argue that a state achieving a strategic monopoly over SuperintelligenceConceptSuperintelligenceAI systems with cognitive abilities vastly exceeding human intelligenceQuality: 92/100 poses a severe national security threat to all other states, analogous to one power gaining unchallenged control of nuclear weapons. MAIM is intended to prevent this outcome by making the attempt prohibitively costly.¹
• Loss of control from unchecked development: Rapid autonomous AI research — what the paper terms "intelligence recursion" — could produce systems that neither their creators nor rival states can effectively constrain. MAIM is intended to slow development to a pace that allows alignment and oversight work to keep pace.¹
• Great-power AI conflict: Unmanaged competition over AI capabilities could trigger direct conflict between major powers. The framework aims to provide a structured alternative to open conflict by channeling competition through sabotage threats rather than direct military confrontation.
• Rogue state and non-state actor acquisition: The nonproliferation pillar of the framework specifically addresses the risk of weaponizable AI capabilities reaching actors outside the deterrence equilibrium — such as non-state actors or states without the capacity for deterrent reciprocity. This relies on mechanisms like US AI Chip Export ControlsPolicyUS AI Chip Export ControlsComprehensive empirical analysis finds US chip export controls provide 1-3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer ...Quality: 73/100 and chip tracking via Compute GovernanceConceptCompute GovernanceThis is a comprehensive overview of U.S. AI chip export controls policy, documenting the evolution from blanket restrictions to case-by-case licensing while highlighting significant enforcement cha...Quality: 58/100 frameworks.

The Three Pillars

The SuperintelligenceConceptSuperintelligenceAI systems with cognitive abilities vastly exceeding human intelligenceQuality: 92/100 Strategy paper presents MAIM within a broader strategic framework:

The deterrence pillar is most novel and controversial. The nonproliferation and competitiveness pillars build on existing policy proposals around Compute GovernanceConceptCompute GovernanceThis is a comprehensive overview of U.S. AI chip export controls policy, documenting the evolution from blanket restrictions to case-by-case licensing while highlighting significant enforcement cha...Quality: 58/100, US AI Chip Export ControlsPolicyUS AI Chip Export ControlsComprehensive empirical analysis finds US chip export controls provide 1-3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer ...Quality: 73/100, and Hardware-Enabled GovernanceApproachHardware-Enabled GovernanceRAND analysis identifies attestation-based licensing as most feasible hardware-enabled governance mechanism with 5-10 year timeline, while 100,000+ export-controlled GPUs were smuggled to China in ...Quality: 70/100.

How MAIM Works

Escalation Ladder

The framework outlines a graduated set of responses to a rival's destabilizing AI development:

The authors argue that states would act preemptively to disable threatening projects rather than waiting for a rival to weaponize a superintelligent system. Whether sabotage at levels 2–4 would be technically achievable is disputed: Google DeepMindOrganizationGoogle DeepMindComprehensive overview of DeepMind's history, achievements (AlphaGo, AlphaFold with 200M+ protein structures), and 2023 merger with Google Brain. Documents racing dynamics with OpenAI and new Front...Quality: 37/100 argues that distributed cloud computing, decentralized training, and adversaries' hardening of AI infrastructure would make effective sabotage difficult short of actions that risk nuclear escalation.³ MIRI notes that the Stuxnet precedent — often cited in support of cyber sabotage's feasibility — affected approximately 11% of intended Iranian centrifuges, and that AI systems can be restored from prior checkpoints after sabotage attempts.⁴

Deterrence Logic

flowchart TD
  A[State A pursues<br/>AI dominance] --> B{Rivals detect<br/>destabilizing project?}
  B -->|Yes| C[Rivals assess threat level]
  B -->|No| D[Project proceeds<br/>undetected]
  C --> E{Credible threat<br/>to balance of power?}
  E -->|Yes| F[Escalation ladder<br/>activated]
  E -->|No| G[Monitoring<br/>continues]
  F --> H[Project sabotaged<br/>or delayed]
  H --> I[Deterrence reinforced:<br/>MAIM equilibrium]
  D --> J[Observability<br/>failure]

  style A fill:#f9d71c,stroke:#333
  style H fill:#ff6b6b,stroke:#333
  style J fill:#ff6b6b,stroke:#333
  style I fill:#90EE90,stroke:#333

The diagram illustrates two pathways from the proponents' model: successful deterrence and observability failure. Critics note that additional failure modes — including false-positive sabotage (misidentifying a non-threatening project as destabilizing) and escalation spirals (sabotage triggering countermeasures rather than renewed deterrence) — are not shown, and that these paths may be as likely as the depicted equilibrium outcome.⁵³

The stabilizing logic requires three conditions: rivals must be able to observe destabilizing projects, they must have credible means to sabotage them, and the threat of sabotage must outweigh the expected gains from pursuing dominance. All three conditions are disputed in the critical literature reviewed below.

Proposed Stabilization Policies

The paper recommends several policies to strengthen MAIM stability:

• Clarify escalation ladders: Establish common knowledge about maiming readiness to prevent misinterpretation of rival actions
• Prevent chip smuggling: Keep decisions about AI development with rational state actors rather than rogue regimes. CNAS (2025) reported an estimated 10,000 to several hundred thousand smuggled AI chips reaching China in 2024, with a median estimate of ~140,000 — evidence that the current nonproliferation regime is already under pressure.⁶
• Remote datacenter placement: Follow the "city avoidance" principle from nuclear doctrine, placing datacenters away from urban areas to limit civilian collateral damage from potential strikes. Hendrycks elaborated in a March 2025 interview that this mirrors how Cold War nuclear doctrine sought to distinguish military targets from population centers — reducing the risk of sabotage escalating into broader civilian harm.⁷
• Transparency and verification: Mutual inspection regimes to reduce false-positive sabotage attacks
• AI-assisted inspections: Deploy "confidentiality-preserving AI verifiers" that can confirm compliance without revealing proprietary details

Differences from Nuclear MAD

While Superintelligence Strategy draws a pedagogical parallel between MAIM and MAD, the authors acknowledge these are structurally different frameworks:

Hendrycks clarified in a follow-up response to critics that the analogy was pedagogical rather than structural, and that the MAIM argument is intended to stand on its own merits, independent of how closely it mirrors MAD.⁸

Major Critiques

Observability Problem

A critique published on AI Frontiers (Arnold, Virginia Commonwealth University, August 2025) highlights that MAIM hinges on nations observing one another's progress toward superintelligence.⁵ AI development happens behind closed doors with breakthroughs often concealed as proprietary secrets. This creates two dangerous failure modes: missing important signs of advancement, or misinterpreting normal activity as a threat and triggering unnecessary sabotage.

Arnold further notes that DeepSeekOrganizationDeepSeekDeepSeek is a Chinese AI research lab founded in 2023 as a subsidiary of High-Flyer, a quantitative trading firm based in Hangzhou. The company gained international attention in January 2025 with D...-R1 demonstrated that algorithmic innovations can drive significant capability gains with a fraction of the compute that infrastructure-monitoring approaches would flag — undermining monitoring frameworks based on compute proxies. Chinese government oversight also provides asymmetric counterintelligence advantages over decentralized Western labs: centralized state control makes it comparatively easier to conceal domestic AI progress from foreign intelligence services.⁵

MIRI's Formal Analysis

Machine Intelligence Research InstituteOrganizationMachine Intelligence Research InstituteComprehensive organizational history documenting MIRI's trajectory from pioneering AI safety research (2000-2020) to policy advocacy after acknowledging research failure, with detailed financial da...Quality: 50/100 published a detailed analysis applying formal deterrence theory to MAIM (Abecassis, April 2025), finding:⁴

• Unclear red lines: What constitutes a "destabilizing AI project" is ambiguous and difficult to monitor; the paper's proposed red line of "fully automated AI research" arrives too late, with minimal response time before a potentially decisive capability gain
• Questionable credibility: Sabotage likely only delays rather than denies rival capabilities — AI systems can be restored from prior checkpoints, weakening the deterrent
• Timing problems: Intelligence recursion might proceed too quickly to be identified and responded to before a decisive advantage is achieved
• Volatile calculus: Immense stakes and uncertainty make deterrence calculations unpredictable

MIRI proposed an alternative regime centered on earlier, more monitorable red lines — such as controls at the advanced chip fabrication stage rather than during active AI development.

RAND Assessment

RAND noted (Rehman, Mueller, Mazarr, March 2025) that the paper makes a contribution to the AI policy debate, but raised concerns that the MAIM world described is "entirely inconsistent with the current reality of private sector-driven AI development."³ The gap between the state-centric deterrence model and the actual landscape of private AI labs raises implementation questions. RAND also argued that MAIM would exacerbate rather than dampen instability by creating potent first-strike incentives — rational actors might calculate it is better to act first rather than wait for a rival to achieve dominance.

RAND additionally noted that China finds current US chip restrictions more deeply threatening than American policymakers may recognize — a perception asymmetry that could accelerate the dynamic MAIM is meant to stabilize.³

Escalation and Moral Hazard Concerns

Additional critiques include:

• Escalation risk: Strikes on AI infrastructure could be perceived as acts of war, since AI infrastructure is deeply intertwined with economic and military power
• Moral hazard: Accepting AI malfunction as a strategic tool could lower ethical standards and reduce investment in proactive safety measures
• Asymmetric perceptions: China may view US chip restrictions as more threatening than American policymakers realize, undermining stable deterrence
• Attribution challenges: Difficulty attributing cyberattacks creates risk of miscalculation and overreaction
• Competing unilateralist approaches: AEI and others argue the US should pursue AI dominance rather than mutual deterrence — that a world where the US reaches superintelligence first is preferable to a MAIM equilibrium⁹

IAPS Stability Analysis

The Institute for AI Policy and StrategyOrganizationInstitute for AI Policy and StrategyAI policy think tank founded in 2024, headquartered in the Washington, DC metro area. Led by Jennifer Marron (Executive Director), Brianna Rosen (Research Director, Frontier Security), Amanda El-Da... published a detailed analysis (Delaney, December 2025) assessing whether a MAIM regime could remain stable long enough for superintelligence to be developed safely.² Delaney used probabilistic reasoning to estimate conditions for "descriptive MAIM":

• ~70% probability China would anticipate near-total loss of great-power status if the US achieves superintelligence unilaterally
• ~60% probability China would pursue escalating sabotage actions (from cyberattacks to conventional strikes) in response
• ~60% probability the US would negotiate a mutual slowdown rather than risk nuclear confrontation

These estimates yield approximately 25% combined probability that all descriptive MAIM conditions are met. Delaney recommended that actors develop verification technologies, advance alignment research, and elevate AI risk awareness in government regardless of which strategic framework ultimately proves correct.²

Limitations

In addition to the external critiques above, MAIM faces structural limitations that constrain what the framework can achieve even if its core deterrence logic holds:

Private-sector AI development: Leading AI labs (including OpenAIOrganizationOpenAIComprehensive organizational profile of OpenAI documenting evolution from 2015 non-profit to Public Benefit Corporation, with detailed analysis of governance crisis, 2024-2025 ownership restructuri...Quality: 62/100, AnthropicOrganizationAnthropicComprehensive reference page on Anthropic covering financials ($380B valuation, $14B ARR at Series G growing to $19B by March 2026), safety research (Constitutional AI, mechanistic interpretability...Quality: 74/100, and Google DeepMind) are private companies, not state agencies. MAIM assumes governments can control the pace and character of their domestic AI development and commit credibly to restraint. Without mechanisms to bind private actors to state deterrence commitments, the framework may stabilize state-to-state relations while leaving private-sector development dynamics unaddressed. RAND described this as the framework being "entirely inconsistent with the current reality of private sector-driven AI development."³

Attribution asymmetry: Even if a sabotage event occurs, definitively attributing it to a specific state actor is difficult in the cyber and covert operations domain. False attribution creates miscalculation risk; correct attribution creates escalation pressure regardless. The problem is compounded by the possibility that non-state actors could conduct sabotage operations mimicking state-level capabilities.

Window of vulnerability during rapid capability gains: If an AI system undergoes intelligence recursion — rapidly accelerating self-improvement — the window for detection and response may be shorter than any deterrence mechanism can accommodate. Fast-takeoff scenarios could occur in timelines incompatible with the espionage-verify-respond cycle MAIM relies on.⁵

Verification bootstrapping problem: The stable verification regimes MAIM requires depend on prior bilateral cooperation to establish. Verification infrastructure requires both parties to agree on inspection protocols, share sufficient technical information, and trust the process — a level of cooperation that deterrence frameworks are generally proposed as an alternative to. As of early 2026, US-China intergovernmental AI safety dialogue has not met since May 2024.¹⁰

Compute-light capability pathways: The DeepSeek-R1 model demonstrated that algorithmic innovations can drive substantial capability gains with a fraction of the compute that infrastructure-monitoring approaches would flag. If the threshold for a destabilizing AI capability can be crossed without building large, detectable datacenters, then monitoring proxies based on compute alone may be insufficient to trigger deterrence responses.⁵

Cooperative Pathways and Skeptical Views

Proponent Arguments for MAIM as Cooperation Scaffolding

Proponents argue MAIM is not merely a framework for mutual threat but a potential pathway toward structured cooperation:

• States may prefer mutual visibility and leverage over an unregulated race, making verification agreements self-interested rather than altruistic
• MAIM's proposed tools — escalation ladders, transparency mechanisms, and verification regimes — could provide scaffolding for legitimately enforceable international agreements over time
• Deterrence could begin with unilateral capabilities and mature into systems of international verification analogous to nuclear arms control treaties
• The framework could eventually evolve into something resembling International Compute RegimesConceptInternational Compute RegimesComprehensive analysis of international AI compute governance finds 10-25% chance of meaningful regimes by 2035, but potential for 30-60% reduction in racing dynamics if achieved. First binding tre...Quality: 67/100 for AI governance

Skeptical Perspectives on the Cooperative Pathway

Critics challenge the assumption that MAIM can serve as scaffolding for cooperation:

• Adversarial entrenchment: Some analysts argue that framing AI governance through deterrence logic entrenches adversarial dynamics, making the trust-building necessary for cooperative governance harder to achieve. Signaling willingness to sabotage rivals may harden strategic calculations rather than opening space for negotiation.
• US policy posture: The Trump administration's AI Action Plan (July 2025) framed US AI strategy as "winning the race" and characterized international governance efforts by international bodies as advocating "burdensome regulations" subject to CCP influence — a posture that treats international agreements as competitive liabilities rather than stability tools.¹⁰
• Dialogue gap: US-China intergovernmental AI safety dialogue has not met since the Geneva meeting in May 2024, even as the Trump-Xi APEC meeting (October 2025) produced an agreement to "consider cooperation on AI" in 2026. The gap between diplomatic signals and operational cooperation frameworks remains substantial.¹⁰
• Private-sector exclusion: If key AI developers are private firms, interstate verification regimes may not bind the actors most likely to trigger a destabilizing breakthrough.

Relationship to Other Frameworks

MAIM intersects with several existing governance proposals:

• Compute GovernanceConceptCompute GovernanceThis is a comprehensive overview of U.S. AI chip export controls policy, documenting the evolution from blanket restrictions to case-by-case licensing while highlighting significant enforcement cha...Quality: 58/100: The nonproliferation pillar relies on controlling access to AI-relevant compute
• US AI Chip Export ControlsPolicyUS AI Chip Export ControlsComprehensive empirical analysis finds US chip export controls provide 1-3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer ...Quality: 73/100: Current US chip restrictions to China are a precursor to MAIM-style nonproliferation
• International Coordination MechanismsConceptInternational Coordination MechanismsComprehensive analysis of international AI coordination mechanisms shows growing but limited progress: 11-country AI Safety Institute network with ~$200M budget expanding to include India; Council ...Quality: 91/100: MAIM's escalation ladders and verification proposals complement international coordination efforts
• AI Development Racing DynamicsRiskAI Development Racing DynamicsRacing dynamics analysis shows competitive pressure has shortened safety evaluation timelines by 40-60% since ChatGPT's launch, with commercial labs reducing safety work from 12 weeks to 4-6 weeks....Quality: 72/100: MAIM attempts to address the same competitive pressures that drive AI Development Racing DynamicsRiskAI Development Racing DynamicsRacing dynamics analysis shows competitive pressure has shortened safety evaluation timelines by 40-60% since ChatGPT's launch, with commercial labs reducing safety work from 12 weeks to 4-6 weeks....Quality: 72/100
• Pause AdvocacyApproachPause AdvocacyComprehensive analysis of pause advocacy as an AI safety intervention, estimating 15-40% probability of meaningful policy implementation by 2030 with potential to provide 2-5 years of additional sa...Quality: 91/100: Some critics argue that advocating for compute pauses or development moratoria would be more effective than deterrence

Developments Since Publication

Since the initial release of Superintelligence Strategy (March 2025), the MAIM framework has generated significant academic and policy engagement:

Follow-up analyses: MIRI (Abecassis, April 2025) and AI Frontiers (Arnold, August 2025) published formal critiques focusing on deterrence theory and observability respectively; IAPS (Delaney, December 2025) conducted the most comprehensive probabilistic assessment, estimating ~25% probability that MAIM conditions are met.⁴⁵² Hendrycks published a defense of the framework on AI Frontiers in 2025, responding to critics including Wildeford and Delaney.⁸

AI-enabled cyber operations: In November 2025, Anthropic disclosed what it assessed with high confidence as a Chinese state-sponsored group (designated GTG-1002) that used Claude Code to execute a large-scale cyber espionage campaign against approximately 30 global targets — operating with AI autonomy at 80–90% of decision points and at speeds "impossible for human hackers."¹¹ The disclosure provided concrete data for the escalation ladder's lower rungs: AI-enabled cyber operations against technology infrastructure are already occurring, though the documented campaign was espionage-oriented (data exfiltration) rather than capability-denial oriented (sabotage of AI development).

AI chip smuggling: CNAS (2025) documented an estimated 10,000 to several hundred thousand AI chips smuggled to China in 2024, with a median estimate of ~140,000 — suggesting the nonproliferation pillar's compute-monitoring assumptions face significant practical challenges.⁶ Of 22 notable PRC-developed AI models identified by 2025, only 2 were trained with Chinese chips, indicating continued reliance on restricted hardware.

US policy trajectory: The Trump administration's AI Action Plan (July 2025) explicitly framed US AI strategy as "winning the race" rather than mutual deterrence, diverging from the MAIM framework's prescriptive recommendations. The paper was released as Trump dropped Biden-era AI safety guidelines.¹²

US-China dialogue: Trump and Xi agreed at the October 2025 APEC summit to consider AI cooperation in 2026, and Trump at the 2025 UN General Assembly called for international verification of AI applications to bioweapons. Some analysts read this as consistent with MAIM-style verification concepts, though no formal verification framework has been proposed and the bilateral AI safety dialogue has not resumed since May 2024.¹⁰

Authors' position: As of early 2026, Hendrycks, Schmidt, and Wang have not retracted or substantially revised the MAIM framework. The revised "Expert Version" (arXiv v2, April 14, 2025) represents the last tracked update to the paper.¹⁸

Key Uncertainties

External Links

• Hendrycks, Schmidt, Wang. Superintelligence Strategy (arXiv:2503.05628, March–April 2025)
• Hendrycks. "AI Deterrence Is Our Best Option" (AI Frontiers, 2025) — proponent response to critics
• nationalsecurity.ai — official paper website