Independent inquiry informed by experience in cybersecurity, software engineering, autonomous systems, and applied AI.
Independent AI Research • Security • Decision Systems
Independent AI research into signals, security, and decisions.
Evening Star AI investigates the places where AI gets brittle: weak signals, adversarial input, uncertain evidence, and decisions that still have to land with a person.
ResearchAnomaly detection, drift, detector agreement, and uncertainty a reviewer can inspect.
Applied LabsExperimental systems for studying security, vulnerability, market, and communications workflows.
PublicationsPublic technical papers on AI risk, evaluation, uncertainty, and governed decisions.
Institute Platform
A founder-led research initiative for applied AI, security, and decision systems.
Evening Star AI is built for a narrow kind of work: publish the research, build the lab system, and test whether the idea holds up when the data is noisy and somebody still needs a decision.
We keep coming back to the same pattern: teams have more telemetry, more scanners, more dashboards, and now more AI output, but the decision still lands on a human trying to figure out what changed, whether it matters, and what should happen next.
The work is strongest when it stays close to real systems: prompts that can be attacked, scanners that over-alert, markets that move before the story catches up, and operators who need a decision they can defend.
About
Why the institute exists
Independent research, experimental prototypes, and technical publications on AI security and decision systems.
Open About ResearchThe technical agenda
Anomaly intelligence, LLM security, vulnerability prioritization, AI-native software assurance, and bounded automation.
Open Research PublicationsWhitepapers and briefs
Technical papers and strategic notes written to be read, challenged, and put to work.
Open Publications LabsWhere the ideas get built
Purple Radar, Purple Firefish, Candles Edge, and related lab experiments connect research questions to testable systems.
Open LabsResearch Scope
Independent research and experimental software.
Evening Star AI publishes independent research and experimental software. It does not offer consulting, freelance development, or client implementation services.
Lab systems test ideas against repositories, prompts, vulnerabilities, markets, and decision records.
AI security, anomaly intelligence, vulnerability intelligence, and decision systems.
Fellowships
Fellowship Program
Evening Star AI recognizes people who contribute serious research, technical review, applied systems, and field experience.
For new contributors, reviewers, students, builders, and technical writers.
For serious researchers, engineers, operators, and applied AI builders leading meaningful work.
For established experts helping shape research direction, technical reviews, publications, and experiments.
Current Research Programs
Current research threads.
The research stays close to places where a wrong answer creates real work for someone else.
Anomaly Intelligence
Weak signals before obvious failure
Unsupervised detection, Isolation Forest, drift, and early warning before the incident has a clean label.
Open Program Adversarial AI SecurityLLM threat detection and testing
Prompt injection, jailbreak pressure, hostile context, risky tool use, and LLM security testing.
Open Program Vulnerability IntelligenceRisk-prioritized remediation
Exposure, exploitability, asset context, scanner confidence, and the remediation decision.
Open Program Decision SystemsConfidence, severity, and action
Recommendations that show confidence, severity, reasons, and what would change the call.
Open ProgramResearch Library
Latest Publications
A short library of papers on AI security, anomaly intelligence, agentic systems, and decisions that need evidence.
The Evening Star AI Engine
The core architecture behind our anomaly work: baselines, detector ensembles, drift, attribution, confidence, and action boundaries.
Red Teaming Agentic AI Systems
A test model for how tool-using agents fail: prompt injection, context poisoning, unsafe tools, permission bypass, and goal drift.
Beacon by Evening Star AI
Give your AI direct access to Evening Star research.
Beacon is Evening Star AI's public Model Context Protocol interface for governed research discovery and retrieval. Compatible AI assistants can discover, search, and read the approved research catalog.
Current Research access requires no credentials and is subject to Beacon policy, rate and concurrency limits, and auditing.
Make AI show its work before it acts.
Useful systems need to show what changed, why it matters, what evidence supports the call, and where a human should slow down or step in.
Evening Star Principle
Useful AI turns weak signals into decisions people can trust.
Discuss the Research
Questions, corrections, and reproducible results help improve the work.
Share a question about a publication, a technical correction, or an idea for research collaboration in AI security, vulnerability intelligence, anomaly detection, or governed decision systems.