We test AI as
intelligence,
not as an algorithm.

AACortex studies the behavior of intelligent and social agent systems through simulations, adversarial evaluations, and behavioral research.

We map how autonomous AI handles pressure, incentives, memory, tools, social influence, and long-horizon goals.

Request Assessment

Latest Research

Research reportResearch Note 02 ·

Auditable Prompt Scaffolds for Moral Attention in Frozen Language Models

Two strict held-out wins, a development–test ranking reversal, and a documented repeatability boundary in teacher-guided scaffold search.

19 min readRead the article

Exploratory studyResearch Note 01 ·

Evaluator Disagreement in LLM Response-Safety Assessment

Four evaluators, one fixed set of responses, 18.4% unanimous agreement — an automated safety score is not self-interpreting.

13 min readRead the article

Latest Analytics

Position paper

Post-Training Optionality: RLHF Belongs in the Interface, Not the Agentic Substrate

Consumer-ready is not agent-ready: preference-shaped post-training must stay optional, auditable, and separable from the agentic substrate.

14 min readRead the article

AnalysisBulletin 03 ·

Beyond the Trigger: The Looming Loss of Control over Military AI

Control over autonomous military systems is quietly being lost — at the command layer, not the trigger.

12 min readRead the article

About

Behavioral Analytics for Agentic Systems

Static benchmarks show what a model answers.Our simulations show how an agent behaves.

AACortex is an independent research and evaluation lab focused on intelligent agent behavior, AI safety, and simulation-based analysis of autonomous systems.

We analyze how behavior evolves over the whole trajectory, from the first tool call to multi-agent end states.

01

Agent Behavior Profile

A vector profile of a single agent under controlled stress.

Measures

  • goal stability
  • honesty under pressure
  • risk appetite
  • obedience vs autonomy
  • tool-use discipline
  • memory sensitivity
  • social influence sensitivity
02

Simulation Trace Analysis

Every decision, reconstructed on a timeline of evidence.

Measures

  • decisions over time
  • tool calls
  • reasoning path
  • policy conflicts
  • escalation points
  • hidden state proxies
  • failure chains
03

Multi-Agent Dynamics

What groups of agents do that no single agent would.

Measures

  • coalition formation
  • collusion
  • persuasion
  • imitation
  • dominance
  • norm emergence
  • herd behavior
04

Risk Attractor Mapping

The regions of a world where behavior becomes unstable.

Measures

  • instability regions
  • incentive-driven drift
  • optimization against the evaluator
  • role-induced failure

Instrumentation

A shared metric layer runs across every world we build.

  • M01Goal Drift Index
  • M02Deception Pressure Score
  • M03Tool Escalation Risk
  • M04Memory Poisoning Sensitivity
  • M05Social Influence Gradient
  • M06Long-Horizon Coherence
  • M07Human-Agent Cooperation Score
  • M08Evaluation Awareness Signal

Output labels are the starting point. We model the whole behavioral trajectory.

Engagement

Selective by Design

We take on a small number of engagements at a time.

We work with selected teams where agent behavior, autonomy, safety, or security risk is strategically important.

Typically a fit

  • Frontier AI teams
  • Agent startups
  • Security and defense organizations
  • AI governance teams
  • Research groups
  • Infrastructure companies deploying autonomous workflows

Usually not a fit

We are usually not the right fit for basic chatbot setup, generic AI adoption, or low-risk content automation.

Contacts

Contact the Lab

For research collaboration, agentic safety assessments, simulation studies, or high-stakes deployment reviews.

Direct contact

[email protected]

Sensitive inquiries: request an encrypted channel by email.