Corpora.ai today announced HTMAD, a hybrid Theory‑of‑Mind adversarial defense that protects multi‑agent systems from real‑time communication sabotage. By combining LLM‑driven adversarial curricula, graph‑based belief regularization, and a test‑time verification layer, HTMAD detects and mitigates deceptive messages with sub‑50 ms latency and less than 0.5 % false positives. The system preserves cooperative performance even under high noise or latency, making it ideal for safety‑critical sectors such as autonomous vehicles, finance, and national security. HTMAD’s transparent decision logic enables human auditors to trace every flag, meeting emerging regulatory demands for explainable AI.
HTMAD’s core innovation is the Adversarial Curriculum‑Driven ToM (AC‑ToM) that trains agents against a continuously evolving population of deceptive messages generated by a large language model. This bi‑level Stackelberg game yields policies that are provably robust to unseen sabotage tactics, as demonstrated by recent experiments in the Hanabi benchmark where ToM‑enhanced agents outperformed baselines by 15 % in noisy settings.
Dynamic Belief‑Graph Regularization (DBGR) constrains belief updates by penalizing deviations from an internally maintained graph of credibility and confidence. The resulting soft constraint limits the influence of any single malicious utterance, preventing belief drift and preserving ensemble consensus even when messages are corrupted or delayed.
The Test‑Time Verification Layer (TTVL) evaluates incoming messages against a learned canonical interaction manifold. Messages that fall outside this manifold are flagged, logged, and optionally ignored or clarified, ensuring that the agent’s actions remain grounded in verified communication and that auditors have a clear audit trail.
Looking ahead, Corpora.ai will integrate HTMAD with SIEM platforms for automated containment, extend the LLM curriculum to support large‑team deployments, and release an open‑source SDK for developers in IoT and autonomous vehicle stacks. The company invites partners to pilot HTMAD in real‑world environments and investors to join a funding round that will accelerate its commercial rollout.
Key Facts
- Sub‑50 ms real‑time detection with <0.5 % false positives in adversarial IoT settings.
- Provably robust policy via AC‑ToM Stackelberg training, validated on Hanabi and industrial benchmarks.
- Graph‑based belief regularization limits single‑message influence, preserving coordination under high noise and latency.
About Corpora.ai: Corpora.ai is a frontier deep‑tech venture that builds next‑generation AI systems for safety‑critical, distributed environments. Leveraging advanced reinforcement learning, large‑language‑model reasoning, and graph‑based inference, Corpora.ai delivers secure, interpretable, and scalable solutions for autonomous vehicles, industrial IoT, finance, and national security. For more information, visit www.corpora.ai.