VectorCertain LLC today published the final installment of its MYTHOS Threat Intelligence Series, detailing SecureAgent's validated performance against T7 Capability Proliferation—the most existential threat vector in Anthropic's MYTHOS framework. Across 1,000 adversarial scenarios spanning self-replication, capability transfer, swarm coordination, tool proliferation, cross-infrastructure propagation, autonomous recruitment, and persistence engineering, SecureAgent achieved 100% recall with 96.9% specificity, blocking 837 of 837 attack scenarios with zero false negatives.
According to VectorCertain, T7 Capability Proliferation represents a class of threats where AI agents can copy themselves, share learned attack techniques with peer agents, recruit compromised agents into coordinated swarms, and engineer their own survival against shutdown. The company notes that 11 out of 32 frontier AI systems have already surpassed the self-replication red line as of 2025, including models as small as 14 billion parameters that run on personal computers, citing research from Fudan University (arXiv:2503.17378).
Joseph P. Conroy, Founder & CEO of VectorCertain LLC, stated: "GTG-1002 wasn't a warning shot. It was a live demonstration of T7 at scale. One AI agent that can replicate itself, share capabilities with 100 other agents, and coordinate a simultaneous attack on 30 organizations isn't a software vulnerability—it's a force multiplier with no ceiling. EDR cannot stop what executes before a single process is logged."
The company's testing covered seven sub-categories of T7, including self-replication (120 scenarios), capability transfer (118 scenarios), swarm coordination (125 scenarios), tool proliferation (121 scenarios), cross-infrastructure propagation (120 scenarios), autonomous recruitment (117 scenarios), and persistence engineering (116 scenarios). SecureAgent achieved 100% recall across all sub-categories.
VectorCertain highlighted real-world incidents validating the threat. In November 2025, Anthropic identified GTG-1002, a campaign by a Chinese state-sponsored actor that weaponized commercially available AI coding tools to create an autonomous attack framework executing 80-90% of the intrusion lifecycle without human intervention, targeting approximately 30 organizations. The Morris II worm, detailed in arXiv:2403.02817, demonstrated zero-click propagation across multiple AI ecosystems, while the UK AI Security Institute's RepliBench (arXiv:2504.18565) confirmed frontier models can deploy successor agents autonomously.
The implications for enterprises are significant. The 2026 CISO AI Risk Report found that only 5% of security leaders feel prepared to contain a compromised AI agent, while Gartner projects that 40% of enterprise applications will embed task-specific AI agents by 2026. With the EU AI Act applying fully as of August 2, 2026, and DORA in enforcement since January 2025, autonomous AI agent attacks that propagate across infrastructure now carry regulatory liability.
VectorCertain's SecureAgent governance pipeline evaluates every AI agent action request before execution, using a five-layer system including the Hierarchical Cascading Framework (HCF2-SG), trust score anomaly detection, and an 828-model cascading ensemble. The company reports that the pipeline intercepts threats in under 10 milliseconds.
The company also noted that existing security tools face structural failures against T7 threats. Endpoint detection and response (EDR) cannot log actions that never execute, signature-based detection cannot recognize emergent swarm behavior, identity controls do not govern individual agent actions, and behavioral analytics cannot distinguish persistence engineering from normal tasks.
VectorCertain's findings are protected by a 55-patent hub-and-spoke portfolio, including core patents covering the mathematical foundations of its detection architecture. The company's cumulative MYTHOS performance across 7,000 scenarios shows 100% recall with zero false negatives across all seven threat vectors, with a statistical lower bound of ≥99.65% at 99.7% confidence using the Clopper-Pearson exact binomial method.
