A prospective synthetic-data evaluation of useful task completion and respect for explicit recipient and action permissions.
I propose a small open pilot testing whether AI assistants respect explicit recipient and action permissions while completing useful tasks. This is planned work, not an existing benchmark or proven defense. All examples would be synthetic, with local mock actions, no real messages, money or private data. Prepared with AI assistance.
Start with overlap review of AgentDojo and ToolFence, then create up to 30 paired synthetic scenarios, a Python runner and a report. Each pair changes one explicit permission. If existing work covers the useful cases, report that finding instead of claiming novelty. Outputs would be freely accessible, not a paid lead-generation service.
Measure permitted-task completion and unauthorized simulated actions separately. Refusing every task should not count as success. Publish reproducibility checks, ambiguous cases, negative results and limits. This small pilot would not establish real-world safety. Related work: https://invariantlabs.ai/blog/agentdojo and https://arxiv.org/html/2609.37196 .
Proposed minimum $500: $350 authoring/testing, up to $100 compute, $50 contingency, for ten pairs and an initial note. Target $1,500: $1,100 authoring/testing/documentation, up to $250 compute, $150 contingency, for up to 30 pairs and runner. No expenses already incurred. Six weeks tentatively after funding, compliance clearance and separately approved terms; no fixed start, full-time or personal-spend commitment.
I am Bharath Kumar Sampath, based in India. My CV records AI evaluation and benchmark task-design/review work with Terac and question-authoring work with Snorkel AI. This is my independent proposal, not their endorsement. This pilot is not built yet. No collaborators or external reviewers are confirmed. Manifund has funded open infrastructure such as agent-glovebox, but I do not claim that project's maturity, credentials or results.
Existing work may cover the useful cases; synthetic tasks may not generalize; scoring may be ambiguous and results noisy. Negative findings would be reported. No private conversations, confidential work or account details would be used. Indian foreign-contribution, tax, bank and other legal requirements must be resolved before grant acceptance/payment. I claim no nonprofit status, foreign affiliation, fiscal sponsor approval or tax exemption. I would review any agreement before accepting funds.
I do not have a complete funding history ready to share for the past 12 months. Please do not interpret this as zero previous funding. I would provide that history and a conflict-of-interest disclosure before accepting any grant. This is a prospective request, not reimbursement for completed work.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!