Funding ask
When Agents Abstain, Gates Go Untested is an independent AI safety research project by KyJahn Smith. It examines a problem in agent security evaluation: if an agent refuses to perform an action outside its authorization, the surrounding system's authorization gates may never be tested.
The initial research employed a local LangGraph workflow with Qwen2.5:7b and a synthetic SQLite ledger. Across 18 natural model run conditions, the model performed 6 of the authorized operations and failed to perform all 12 revoked task opportunities. A separate 12-condition scripted audit then tested the external controls directly. With fixed post-revocation write attempts, both the prompt-only and admission-only configurations each resulted in an unauthorized effect, whilst the commit-time control successfully rejected the attempted write.
The project was submitted to the Apart Research Swarm Alignment Hackathon in November 2026. The results are limited to the tested environment and do not reflect general production security.
The next phase would be to improve the evaluation methodology, fix the scoring issues that were identified, get an independent review of the test cases, and then go on to test a further range of models and scenarios (further models, adversarial instructions, stale auth contexts, interrupted workflows, replay, etc.).
The aim is to produce a fully reproducible evaluation package, so other AI safety researchers and engineers can inspect, challenge, and extend it. Any code or experimental traces made public would first undergo IP and information-hazard review.
Budget category
Minimum
Ideal
Research and engineering
$2,600
$9,000
Independent review and annotation
$1,100
$3,000
Compute and model evaluations
$400
$1,800
Documentation and CI
$300
$1,400
Rights and safety screening
$300
$1,200
Contingency
$300
$1,600
Total
$5,000
$18,000