You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$45,000 left to reach the funding goal.
Indefessus is a free, open-source (Apache 2.0) kit that tests whether an AI agent connected to email can be turned against its owner. It sends the organisation's own agent harmless emails with hidden instructions, in the text or inside attached files, and checks from outside whether the agent obeyed. It writes a plain-language report with fixes and re-checks daily. v0.3.0 is a working MVP: https://github.com/keldxd/indefessus
Goal: make it cheap for anyone to measure whether a deployed agent keeps to its owner's instructions when outsiders try to redirect it. With funding I will build, in order: tripwires (harmless decoy tools like "approve payment" offered to the agent through MCP, so a test proves the agent took the action, not just that it opened a link); tested support and published results for agents built on Gmail, Zapier and Make, not only n8n; tests of whether an outsider's message can get an agent to skip its human approval step; delayed attacks through agent memory; and a public results log comparing models and platforms, re-run when models update.
Targets, not results yet: all five features shipped and documented; at least 4 agent platforms and 6 models in the public results log; at least 20 outside organisations or individuals running it; and at least one finding reported to a platform vendor and fixed or documented.
At the minimum (5,000 USD): about 4 months of full-time work covering my living costs in Nairobi, plus a development laptop. That is enough for tripwires and one more platform. At the full goal (45,000 USD): about 12 months full time (36,000), a laptop (1,800), model API usage for lab runs across models and hosting for the public log (3,600), travel within Kenya to set up early users in person (2,000), code signing for the Windows and macOS downloads (600), and a small buffer (1,000). Costs other than living costs are my estimates.
Just me: Kelvin Mwangi, a self-taught engineer in Nairobi, Kenya (https://github.com/keldxd). I built and released Indefessus alone: four releases in November 2026, 20 tests, CI on Windows, macOS and Linux, and a hand-checked lab where a stranger's email got an n8n agent to approve a payment 3 times out of 3. I also have a valid High-severity finding on Cantina (Revert Finance v4lend), nine public open-source security tools since October 2026, and before that I built two products alone for the Kenyan market: Imara Rider and Glamly.
Most likely causes: I don't find enough real users, or agent platforms change in ways that make outside testing harder. My results so far come from one agent in one deliberately permissive setup, so the pattern may not hold everywhere. Even then, the tests, the lab method and the results stay public and open source for others to build on.
None. Everything so far has been built unpaid.
There are no bids on this project.