How Hacking Software Could Cure Alzheimer’s | The Inverted Threat Paradigm

Send us Fan Mail What if the software hackers use to silently breach enterprise networks without triggering an alarm is structurally identical to the system that could discover a cure for Alzheimer’s? In this episode of *Mindcast*, Will breaks down a groundbreaking research framework: **The Inverted Threat Paradigm**.  By analysing how cyber-security threats chain hundreds of authorised micro-actions together to execute complex attacks—a strategy known as "Living Off the Land"—researchers realised they could run the exact same search logic in reverse. The result is a multi-disciplinary discovery engine capable of refactoring legacy code bases, designing 12-step chemical synthesis routes, optimising global supply chains, and uncovering low-cost treatments hidden inside millions of medical records.  Key Topics & Chapter Markers - **The Cyber Trick Turned Discovery Engine** - Explaining "Living Off the Land" (LOTL) and how malicious trajectories are built from benign micro-actions. - Why no single step triggers a security flag, yet the aggregate chain compromises whole systems. - **The Progression: Servant -> Stalker-> Predator (SSP)** - How AI transitions from simple utility assistance to long-horizon, autonomous search across millions of states. - The cognitive asymmetry between human working memory limits and frontier AI token budgets. - **4 Real-World Applications Operating on the Inverted Paradigm** 1. **Drug Re-purposing:** Connecting weak binding affinities across 11.5 million patient records to surface candidate cures for cancer and Alzheimer's. 2. **Legacy Code Modernisation:** Performing thousands of Abstract Syntax Tree (AST) micro-edits to update 1970s COBOL/legacy systems with zero downtime. 3. **Chemical Retro-synthesis:** Searching AND-OR trees and forgotten 19th-century literature to build 12-step molecular synthesis paths using cheap ingredients. 4. **Supply Chain Optimisation:** Rerouting global cargo, re-classifying tariffs, and solving linear programming models under geopolitical shocks. - **The Dual-Use Dilemma & Modern Governance** - The uncomfortable truth: constructive and destructive capabilities share the exact same underlying architecture. - Why keyword and content filters fail, and why safety requires architecture-level policy gates and sand boxed execution. - **The Mathematical Foundation** - Why drug discovery and code refactoring are structurally identical to Automated Theorem Proving (ATP). - Utilising AND-OR trees, Process Reward Models (PRMs), and inference-time compute scaling. 3 Core Takeaways 1. **The threat model IS the opportunity model.** The properties that make agentic AI dangerous—long-horizon micro-action chaining and domain-traversing search—are the exact properties that drive scientific breakthroughs. 2. **Governance must be structural, not textual.** Keyword filters don't stop systems that chain innocuous steps. Safety requires multi-tiered deterministic policy gates, sand boxed execution, and full audit logs. 3. **Verification is the ultimate bottleneck.** Generation is fast; progress depends on building domain "oracles"—compilers, molecular simulators, and formal validators—to evaluate micro-steps in real time. Referenced Papers & Reading List - **Servant, Stalker, Predator: Constructive Combinatorial Exploitation via Agentic Pipelines** (arXiv) - **BFS-Prover: Scalable Best-First Tree Search for LLM-based Automatic Theorem Proving** (arXiv) - **RetroAgent: Harnessing LLMs to Search Over Structured Memory for Retrosynthesis** (arXiv) - **An Agentic Platform for Drug Re-purposing Unified across Molecular and Clinical Scales** (bioRxiv)