Research Programs
Research programs for AI systems that need evidence, controls, and review.
Evening Star AI studies the parts of AI that get risky once models touch tools, code, data,
security operations, and human decisions.
These programs organize independent research: the questions being studied, the experiments being built,
the findings being published, and the limits that still need testing.
The evaluation approach is documented on the methodology page.
These are research programs and experimental systems, not consulting or implementation services.
Current Research Programs
Current research focus.
The work is published as short papers and briefs for builders, security teams, leaders, and operators.
One current thread is the shift from code scanning to release judgment: not just finding issues,
but studying the evidence people need to evaluate whether AI-assisted software is safe enough to ship.
Anomaly Intelligence
Unsupervised AI for detecting weak signals before they become obvious incidents.
The problem
Operational environments rarely have clean labels, so early warning depends on drift, outliers, detector agreement, and interpretable anomaly evidence. This thread came from watching teams collect more signals without getting any closer to a confident decision.
Where it shows up
The Evening Star AI Engine research examines Isolation Forest, robust scoring, drift detection, and weak-signal summaries as a reusable intelligence core.
Where it is going
Richer feature attribution, better detector ensembles, and anomaly memory that can compare current behavior against prior regimes.
Adversarial AI Security
Security research for LLM systems exposed to malicious instructions, jailbreak pressure, and untrusted context.
The problem
LLM applications increasingly connect to tools, files, memory, and workflows, making prompt injection and adversarial inputs operational security problems.
Where it shows up
Purple Firefish explores prompt injection detection, jailbreak detection, adversarial-input scoring, LLM threat modeling, and AI security testing.
Where it is going
Policy-aware gateways that inspect prompts, retrieved content, tool calls, outputs, and agent memory before action is allowed.
Vulnerability Intelligence
Evidence-first prioritization for vulnerability risk, exposure, exploitability, and remediation decisions.
The problem
Security teams do not need another flat list. They need to know what is exposed, exploitable, important, and fixable now.
Where it shows up
Purple Radar research explores how vulnerability prioritization, exposure analysis, risk scoring, and exploitability signals can inform reviewable findings and reports.
Where it is going
Risk models that combine asset context, adversary signal, scanner confidence, known exploitation, and remediation feasibility. This research is especially valuable in environments with stale asset data, fragmented ownership, and unclear approval boundaries.
Decision Systems
Models and interfaces that convert evidence into confidence, severity, explanation, and next-best action.
The problem
A score is not enough. People need reasons, severity, uncertainty, and a practical next step. The hard part is not detecting one strange event; it is deciding whether that event should change what a human does next.
Where it shows up
Evening Star lab experiments explore confidence scoring, severity modeling, reason codes, and explainable recommendations across multiple domains.
Where it is going
Next-best-action engines that produce trigger levels, invalidation logic, human approval points, and decision logs.
Agentic Automation
AI agents that monitor signals, summarize risk, automate bounded workflows, and support human operators.
The problem
Agents become valuable when they can do useful work without hiding evidence, assumptions, handoffs, or accountability. We care less about agent theater than whether the handoff can be inspected.
Where it shows up
Agentic communications work explores how AI can monitor operational exhaust, enrich signals, summarize risk, and prepare reviewable actions.
Where it is going
Human-in-the-loop agents that can watch, brief, recommend, queue action, and document outcomes without overstepping authority.
AI-Native Software Assurance
Research into how software can be trusted when AI participates in design, implementation, review, testing, and release.
The problem
AI coding assistants and engineering agents can speed delivery, but they also introduce insecure logic, false confidence, hallucinated dependencies, unsafe tool use, and unreviewed architectural change.
Where it shows up
This research connects to repository evidence, release review, dependency risk, configuration drift, and human-verifiable decisions for AI-assisted software.
Where it is going
AI-native release review, repository prompt-injection testing, agentic code-change controls, software supply-chain evidence, and release cards a human can challenge.
AI Governance
Practical control patterns for AI systems that need accountability, audit trails, and safe operating boundaries.
The problem
Capability is not the same as control. Deployed AI needs policy boundaries, documented assumptions, approval points, and records that survive review.
Where it shows up
Evening Star lab experiments examine how confidence, severity, reason codes, and decision logs can help people understand and challenge model output.
Where it is going
Governance patterns that work inside agent workflows, AI security gateways, vulnerability programs, evaluations, and audit-ready reporting.
Applied AI Labs
Domain-specific experiments that test Evening Star research against realistic workflow constraints.
The problem
Research gets sharper when experiments account for engineering constraints, messy data, human review, and domain-specific decisions.
Where it shows up
Purple Radar, Purple Firefish, Candles Edge, and related lab experiments connect the research to testable systems.
Where it is going
More domain-specific AI systems that reuse the same anomaly, security, decision, and governance architecture.
