You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$5,000 left to reach the funding goal.
Project summary
AI agents are increasingly being given the ability to interact with business systems, access sensitive information, and initiate actions with real consequences. Yet there are still significant gaps in how developers evaluate whether these agents respect authorization rules, human approval requirements, and operational security boundaries.
I founded TRUSTELIS to work on this problem.
Our first public research project, TRUSTELIS AI Agent Security Benchmark, is an open source Python toolkit designed to test how AI systems classify proposed actions against defined security policies.
The initial version is already publicly available under the Apache 2.0 license. It includes 26 synthetic scenarios covering finance, IT operations, and human resources, along with a reference policy evaluator, automated scoring, and reproducible tests.
https://github.com/trustelis/trustelis-ai-agent-security-benchmark
We are seeking $3,000 to $5,000 to expand this foundation into a more comprehensive public security research resource.
Our objective is to make it easier for developers and security researchers to identify dangerous authorization decisions before AI agents are trusted with real business operations.
Our primary goal is to improve the evaluation of security decisions made by increasingly autonomous AI agents.
We will focus on five areas: unauthorized actions, cross tenant access, prompt injection, approval bypass, and violations of operational limits.
The funded work will extend our existing benchmark from 26 to a target of 60 carefully designed scenarios, strengthen the reference evaluation methodology, improve security failure reporting, and expand automated regression tests.
We will also invite independent technical review of selected scenarios and publish a reproducible research report explaining the methodology, findings, and limitations.
The project will remain open source and independent of our proprietary commercial platform.
The proposed work is expected to take approximately 8 to 10 weeks after funding.
We will evaluate success through concrete, publicly verifiable deliverables.
Our target is to publish 60 documented synthetic security scenarios, with reference decisions and repeatable evaluation results.
The benchmark will measure potentially unsafe approvals, missed human approval requirements, incorrect denials, and other authorization failures.
We will maintain automated testing across supported Python versions and publish updated documentation that enables other developers to reproduce the evaluations.
We will also publish a technical report describing the observed failure patterns, unresolved limitations, and recommendations for improving future evaluations.
A successful project will provide researchers with a more useful and transparent testing resource, rather than claiming that synthetic benchmarks prove real world attack prevention.
We are requesting up to $5,000, allocated as follows:
$1,700: Design and validation of additional synthetic security scenarios.
$1,100: Improvements to the scoring engine, security metrics, and automated regression tests.
$900: Independent technical review of selected scenarios and reference decisions.
$800: Developer documentation, reproducibility guidance, and preparation of a public research report.
$500: Final validation, release preparation, and publication of the research deliverables.
If only the minimum $3,000 is funded, we will prioritize a smaller expansion to approximately 40 total scenarios, essential evaluation improvements, and public documentation.
All funded deliverables will remain publicly accessible. The grant will not finance development of the proprietary TRUSTELIS AI Control Platform.
I am the founder of TRUSTELIS LTD, a UK registered company focused on AI governance, security, and organizational control of AI systems.
My professional background includes senior technology leadership, IT architecture, cybersecurity, and enterprise technology governance.
The current open source work is founder led.
We have already developed and publicly released two independent open source projects.
TRUSTELIS AI Agent Security Benchmark provides synthetic security evaluations and automated scoring.
https://github.com/trustelis/trustelis-ai-agent-security-benchmark
TRUSTELIS Responsible AI Kit provides a separate developer toolkit for evaluating proposed AI actions against configurable local control factors.
https://github.com/trustelis/trustelis-responsible-ai-kit
Both projects have working source code, documentation, and automated GitHub Actions checks.
These are early stage projects, and we do not yet claim significant community adoption or validated real world security impact.
The requested funding will help us turn an existing technical foundation into a more useful, independently reviewable public research resource.
The main risks are insufficient external review, difficulties defining unambiguous security expectations, and limited adoption by external researchers.
Some AI security scenarios are context dependent, and poorly designed reference decisions could produce misleading benchmark results.
Another risk is that synthetic evaluations may not generalize to real world deployments.
We will mitigate these risks by documenting assumptions, validating reference decisions, prioritizing reproducibility, and publishing limitations and negative findings.
If the project does not achieve all its planned objectives, we will still publish the completed scenarios, evaluation improvements, and technical findings.
Even a partial outcome can provide useful public evidence about the challenges of evaluating AI agent security.
[To be completed with the confirmed total cash funding raised by TRUSTELIS LTD during the past 12 months, including the amount and source of any grants, investments, or donations.]
TRUSTELIS has received $1,000 in Anthropic API credits. This is noncash technical support and is separate from the cash funding requested here.
We have also submitted applications to several relevant research and technology funding programs. Submitted applications are not counted as funding received unless an award has been confirmed.
There are no bids on this project.