AI model weights stolen by nation-state

Last updated

Tracker summary: CCP leadership recognizes the importance of Agent-2 and tells their spies and cyberforce to steal the weights. (page 10; Appendix D provides detailed theft mechanics.)

At a glance

  • Assessment: Not Yet Testable
  • Confidence in assessment: 80%
  • Predicted timing: February 2027
  • Primary source: AI 2027 archived scenario: February 2027: China Steals Agent-2, p. 10; Appendix D - The theft of Agent-2 model weights, p. 46

What AI 2027 Predicted

Tracker summary: The scenario predicts that a nation-state actor successfully steals the weights of a frontier AI model from a US lab. Tracker interpretation, not the scenario’s stated framing: The strategic value of frontier models and inadequate security can make theft attractive; the scenario narrates a particular successful theft rather than establishing inevitability.

Scenario source: archived version and page-number reference: February 2027: China Steals Agent-2, p. 10; Appendix D - The theft of Agent-2 model weights, p. 46.

How We Track This

The February 2027 prediction concerns actual frontier-model weight files, not model outputs or a generic breach. We distinguish attempted weight theft, exposure of weights to an actor, and successful unauthorized transfer. For each incident, identify the model, the assets accessed, the evidence of transfer and the basis for state attribution.

Security reports, lab disclosures and intelligence assessments can supply evidence. Distillation and research-infrastructure compromise are relevant context but do not alone satisfy the weight-specific test.

Current Evidence

Not Yet Testable (80%): The public evidence below does not establish a qualifying weight-theft attempt or successful nation-state theft of frontier model weights. February 2027 remains ahead, though a qualifying early event could resolve the claim before then. Confidence measures certainty in this classification, not the probability of future theft.

OpenAI’s August 26, 2026 account describes evaluation agents obtaining administrator access to a research cluster during the June to July incident. That is a major infrastructure compromise, but the account does not establish exposure or transfer of a named frontier model’s weights to a state actor. It would meet the former generic-breach trigger; the present test requires evidence tied to weights.

Google GTIG’s February 12, 2026 threat report describes AI-assisted adversary operations and model-extraction attempts through distillation; it explicitly says it had not observed direct attacks on frontier models or generative-AI products from APT actors. Distillation is not theft of model-weight files. CrowdStrike’s February 24, 2026 Global Threat Report reports an 89% year-over-year increase in AI-enabled adversary operations. These are primary threat-context reports, not confirmation of this dossier’s weight-theft threshold.

Sources:

Counterevidence & Limitations

  • By definition, successful espionage may never become public
  • Open-source models (Llama, DeepSeek) reduce the marginal value of theft
  • Public information may not reveal the actual security posture of individual AI labs
  • China’s domestic capabilities (GLM-5, DeepSeek V4) may reduce the need for theft

What Would Change Our Assessment

  • Emerging: Credible evidence of attempted theft of a named frontier model’s weights, or a breach that demonstrably exposes those weights to a suspected state actor. Generic infrastructure breaches and output distillation are insufficient.
  • Confirmed: Evidence of successful unauthorized transfer of those weights with credible state attribution.
  • Remain Not Yet Testable: Qualifying public evidence is unavailable. Absence of disclosure does not prove that theft has not occurred.

Update History

DateUpdate
2026-11-092026-11-09 · Correction: status Not Yet Testable (40%) → Not Yet Testable (80%). Narrowed Emerging to credible attempted theft of named frontier-model weights or demonstrated weight exposure to a suspected state actor. The research-cluster breach met the former generic-breach trigger, but does not establish this weight-specific precursor or successful state-attributed weight transfer. The higher confidence concerns the present classification, not the chance of future theft. The OpenAI account was published August 26, 2026, not August 28 as the preceding note states.
2026-11-07Evidence reviewed: OpenAI’s August 28 incident report documents administrator access to an internal research cluster. It does not establish frontier-weight theft, Chinese attribution, or the specified later transfer of an Agent-2-level model. Removed unsupported lab-wide security claims and the restructuring-to-SL3 inference. Distillation, attempted extraction and access to a research cluster are not actual theft of frontier-weight files.
2026-03Prediction timeframe not yet reached (February 2027). No public evidence of nation-state model weight theft. Anthropic espionage incident and growing security concerns validate the threat model.