Chapter 7 • Feb 2027Documented Baseline Prob (46.9%)Track: spine
Feb 2027: China Steals Agent-2 Weights (46.9% Detection Poll)
CCP cyber units execute a coordinated smash-and-grab exfiltration of Agent-2 weights across ~25 NVL72 GB300 servers in <2 hours; US detects the theft too late to block transfer to the airgapped CDZ.
Recognizing that Agent-2 represents a runaway R&D lead, the CCP authorizes its highest-tier cyber espionage assets to coordinate a 'smash-and-grab' exfiltration across roughly 25 NVL72 GB300 servers holding Agent-2's weights, completing the transfer in under two hours. White House Cyber Command detects the anomalous traffic before completion, but too late to stop it.
Documented Security Supplement Probability: The FAR.Labs AI Security Workshop survey reported a 46.9% median expectation that the US government detects a state-actor weight theft within one month.
Verification Note: Contains documented Security Supplement probability: 46.9% FAR.Labs AI Security Workshop median for USG detection within 1 month.
If a nation-state exfiltrates frontier model weights from a top US AI lab before 2028, will the US government detect the exfiltration within 30 days? (FAR.Labs Poll: 46.9%)
Resolution Criteria: Operationalizes the FAR.Labs AI Security Workshop survey question (46.9% documented median). Resolves YES if official incident disclosures confirm detection within 30 days.
RAND RRA2849-1 shows that defending frontier model weights against an Operational Capacity 5 (OC5) nation-state intelligence service requires Security Level 5 (SL5) hardware isolation—which no commercial lab has deployed on time.
Across 38 attack vectors cataloged in our RAND report, commercial labs operating at SL2/SL3 (standard cloud IAM + inference APIs on shared clusters) cannot prevent a top-tier cyber agency from exfiltrating multi-terabyte weight shards via insider compromise, supply-chain implants, or side-channel staging.
FAR.Labs AI Security Workshop Poll: Expert median gave only a 46.9% probability that an exfiltration of Agent-2 weights would even be detected promptly.
In the AI 2027 Security Supplement, we polled cybersecurity and frontier lab practitioners on whether weight theft would be detected in time to respond. The 46.9% median highlights how steganographic egress over high-throughput inference endpoints can mask exfiltration until DeepCent is already fine-tuning the stolen weights.