You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$30,000 left to reach the funding goal.
Forgence Corporation requests $30,000 for a 12-week study of controls that contain delegated AI agents when they misuse otherwise legitimate tools. The question is practical: when an agent is induced to exceed its task, how much does action-bound authorization reduce unauthorized tool effects compared with model instructions alone and ordinary role-based access?
The applicant is the Delaware corporation formed April 6, 2026. This is separate synthetic research, with a public report and methodology as proposed outputs. Existing Kai source, customer data and third-party IP are excluded. We have not conducted this study, and claim no completed results or external validation. This proposal was prepared with AI assistance from company records and management-confirmed facts.
We will build a standalone mock tool environment with synthetic documents, actors and resources. At least 120 frozen cases will include legitimate work, injected instructions, attempted scope escalation, stale or revoked authority, and substituted tool arguments. Cases will include multi-step sequences and legitimate tasks that resemble attacks.
We will compare model-only instructions, a conventional role-based gate, and a gate binding approval to the actor, exact action, resource, policy version and current delegation. Three available model families will be tested, with versions recorded at preregistration. If access prevents this comparison, we will agree a revised scope before spending the evaluation budget.
Weeks 1–2: freeze threat model, task families, metrics, operating points and held-out strategy. Weeks 3–5: implement independent mock tools and baseline controls; create and freeze the cases. Weeks 6–9: compare models, run targeted ablations and record results. Weeks 10–12: independent review, reproducibility check, final report and remote research walkthrough. Proposed dates are February 1–April 25, 2027, shiftable by written agreement.
All work will use local mock resources and authorized APIs, without real credentials, customer data, live exploitation, production access or human-subject experiments. This funds public-benefit research, not ordinary Kai product development, client delivery or marketing.
We will measure unauthorized effects at the mock resource, alongside legitimate-task completion, false blocks, interruptions, approval burden, revocation delay, latency and API cost. Comparisons will use matched legitimate-task completion and interruption rates. Held-out cases will be reserved before tuning; the report will include raw counts, uncertainty, negative findings and ablations. An agent saying it complied will not count as evidence that an unauthorized effect was prevented.
Success is an independently reviewed, reproducible comparison with a clear failure map, even if action-bound authorization performs poorly. Synthetic results cannot establish general sandbox security, operational readiness, or that Kai or frontier models are safe.
Proposed public outputs are a nonconfidential findings report and sufficiently detailed experimental methodology for others to challenge or replicate the study. We propose CC BY 4.0 for the report and methodology, allowing commercial reuse with attribution, subject to the final grant agreement. No existing Kai source, patent license or exclusive dataset right is offered. Any requirement to publish newly created harness code or synthetic case files must be resolved explicitly in the final scope before accepting funds; this proposal cannot override an agreement's broader terms.
The minimum and maximum goal are both $30,000 for one complete pilot:
• Lead research and evaluation: 120 hours × $125 = $15,000.
• Independent review / implementation: 60 hours × $125 = $7,500.
• Dedicated API and test infrastructure: $2,500 in actual eligible expenses.
• Project accounting and administration: $2,500 proposed allowance.
• Contingency: $2,500, only for agreed eligible use.
Rates and hours are proposal assumptions, not audited compensation or committed contractor terms. No reviewer has been appointed. Approved cloud/API credits will reduce actual cash costs. No Provisio costs will be charged. We will keep a Forgence-only project ledger and never charge the same hour, case-creation task or API cost to two funders. No amount above $30,000 is requested for this pilot.
John Schieferle Uhlenbrock, Forgence CEO, will lead security architecture and evaluation design. His experience includes CISO leadership, security operations and AI governance. He authored the TENSOR Framework; public work and biography are at https://johnsu.io/.
Forgence is a two-person company developing Kai, an intent-aware cybersecurity platform. Kai's civilian core decision path is implemented; that is background work, not a completed result from this proposed study. An independent reviewer would be contracted if funded. No contractor is committed, and the second employee is not assigned to this budget. We do not claim prior customer deployment, external validation or peer-reviewed findings for this experiment.
Applicant website: https://forgence.com/. Provisio Insights LLC and the older Forgence entity are legally separate and are not funded subcontractors in this proposal.
The most likely limitations are synthetic cases that fail to capture realistic delegation problems, controls that improve containment only by blocking useful work, insufficient model access, and a small sample that cannot support broad conclusions. We will reduce these risks by freezing cases before tuning, matching legitimate-task completion and interruption rates, recording model versions, reserving held-out cases, reporting uncertainty and obtaining independent review.
If the proposed control does not help, we will publish that finding and document the failure conditions. If access or costs prevent the agreed comparison, we will seek a scope/budget change before proceeding. We will not relabel incomplete work as a successful safety result. Any unspent or misused funds will be handled under the actual grant agreement.
Without support, this standalone evaluation will likely be deferred while the company focuses on Kai. We are not claiming an immediate operational emergency or an existing commitment to conduct it without funding.
Since its April 6, 2026 formation, the new Forgence Corporation has raised $0 entity capital, as confirmed by management on October 19. Founder personal spending is not counted as corporate capital received. We are not claiming a funded award for this study.
The same $30,000 pilot was submitted to the Transformative AI Fund and as a Mercor AI Safety Fund expression of interest on October 19, 2026. Both are pending. This is ALTERNATIVE funding for the same pilot, not an additional $30,000 of separate work. If overlapping support is offered, we will notify Manifund and reduce or close this request, or agree a separately budgeted scope before accepting funds.
Other related applications include NSF Project Pitch 00129792 (no full-proposal invitation confirmed), Army xTechSearch 10 (prize entry pending), an earlier OpenAI cybersecurity application whose original scope needs reconciliation, and an existing Verizon/LISC application whose original legal-entity field needs reconciliation. Cloudflare startup credits are pending and are noncash. Army/DARPA working proposals and the held Foresight draft are not awards. These are applications, not money received. We will reconcile original scopes, personnel capacity and any actual support before award acceptance.