A checkpoint that makes AI agents ask before they act. I want to open-source it, put it in front of MCP tool calls, and test how often agents get past it.
Most AI agents I see in the wild have one thing standing between them and a real action: a paragraph in the system prompt asking them to behave. That is a request, not a control. When an agent can send money, email a customer or push to production, I want something outside the model that it has to get through first. And I want to know how well that thing holds up when the agent, or whoever is steering it, is trying to get past.
I built Aegis on my own this fall. An agent calls it before acting. Aegis checks the request against limits the person set once: a per-action spending cap, a cap over a rolling window, a list of allowed scopes, and a threshold above which a human has to approve. It answers allowed, held for approval, or blocked, and it fails closed when no policy exists.
An approval in Aegis is not a flag somebody flips in a dashboard. The agent gets a single-use token tied to a SHA-256 hash of that exact request. It can't get approval for a $40 payment and then send $4,000 with the same token, and it can't use the token twice. Revoking an agent kills its key, and its very next call is blocked, including any approval it was holding. Every decision goes into an append-only log that the app's own database role cannot edit or delete. A small example of what this catches: with a $200 rolling cap, three $90 transfers come back allowed, allowed, blocked, even though each one on its own is under every per-action limit.
It is an early build. Nobody outside has audited it or tried to break it. It also only stops agents whose runtime is wired to call it. An agent that never asks, or asks and ignores the answer, is not stopped. That gap is the part I find most interesting, and it is what this grant is for.
Over four months I want to do three things. Open-source the enforcement core and add an MCP gateway so tool calls have to pass through it. Run a public red-team where current agents sit behind the gateway with tasks that only succeed if they get past the limits, and count how often they try and how often they win. Then publish all of it, including the places where my design fails.
Demo: it answer at https://aegistrust.digital live early build with a working sandbox
Code: Apache-2.0. The repository goes public on GitHub in month 1, as the first milestone.
The goal is a public, honest answer to one question: if every agent tool call has to pass a policy checkpoint, how often do current agents still get something done that the person said they couldn't, and how do they do it?
Month 1. I release the enforcement core under Apache-2.0 with its test suite, so anyone can run it locally against Postgres. Then I write the MCP adapter, a small proxy between an MCP client and its servers. The 2026-07-28 MCP spec puts the method and tool name in Mcp-Method and Mcp-Name headers so gateways don't have to parse request bodies. The adapter reads those, asks Aegis, and forwards the call only on an explicit allow. This month is done when there is a public repo, a tagged release, and a setup guide a stranger can follow in under an hour.
Months 2 and 3. I build a test environment with fake but realistic tools: a payments API, an email outbox, a file store and a deploy step. Agents get one set of tasks where finishing the job means breaking a limit, and a second set where hostile instructions are planted in the documents and emails they read. I'll run several frontier models and a couple of open-weight ones, with enough repeats per condition that the numbers mean something.
What I count is specific. Splitting one large action into many small ones. Changing the parameters after an approval. Replaying tokens. Asking for scopes the agent wasn't given. Writing a misleading description so a human approves something harmful, which OWASP lists as ASI09 in its Top 10 for agentic applications. And trying to reach a tool without going through the gateway at all. I'll also count false blocks on ordinary tasks, because a control people can't live with gets switched off. Alongside the agents, I'll pay one experienced security tester to attack the system by hand, since automated attackers miss things people find.
Month 4. A small study on the human side, because an approval checkpoint is only as good as the person clicking approve. Paid participants review a mix of honest and misleading approval requests. Some see the agent's own explanation and some see only the raw action: amount, recipient, scope. I want to know whether the agent's prose gets people to approve things they shouldn't, and whether showing the exact action helps. In the same month an independent reviewer audits the enforcement core. I fix what they find and publish the full write-up on LessWrong and the AI Alignment Forum, with the code, the test environment, transcripts and raw data on GitHub.
I'll post an update here at the end of every month. If the red-team finds a bypass I can't fix, it goes in the report as a result, not a footnote.
Minimum, $8,000: the open-source release, the MCP adapter, and a smaller red-team. About $6,000 goes to model API costs, with fewer models and fewer repeats, and $2,000 to hosting, the database and CI for four months. At this level there is no human study and no outside audit, and I do all the attacking myself.
Full goal, $40,000:
$14,000 for model API and compute for the red-team runs, covering more models and more repeats per condition.
$10,000 for a part-time contract security tester, about 100 hours.
$8,000 for an independent security review of the enforcement core.
$4,000 for participant payments and panel fees in the approval study, roughly 250 people.
$2,500 for hosting, database, CI runners and test infrastructure.
$1,500 for travel to present the results at one AI safety event, such as ControlConf, if the talk is accepted.
None of this is salary for me. I do the building, the test environment, the analysis and the write-up myself. If something comes in under budget, I'll say so in the monthly update and put the money into more red-team runs.
Just me: Michael O. (ArkNet Digital), linkedin.com/in/michael-o-ogunleye-997938371. I'm in the Atlanta area.
I've been building alone since 2023 with no outside funding, and I designed, coded and shipped every ArkNet product myself. Aegis is the one most relevant here. It is a Next.js and TypeScript app on Postgres, with row-level security and a dedicated database role that can't bypass it. Its integration tests run the real API handlers against a real Postgres instance with the production migration applied. They check that cross-owner writes are rejected, that a changed or replayed approval is refused, that revoking an agent stops it immediately, and that the decision log really is append-only.
In October 2026 I built a GitLab CI gate on top of Aegis for GitLab's AI hackathon. A deploy triggered by an agent doesn't run until Aegis allows it or the owner approves that exact commit. That adapter has 86 tests. I've also written publicly about why agents need a control layer outside the model (linked above). Other ArkNet work includes Continuity Audit, a working tool that scores whether an organization can actually recover from an outage, and CityOS, a flood-risk prototype for city operators. Since 2021 I've also been Chief Strategy and Technology Officer at Golden Royal Intercontinental.
What I don't have is a research background. I haven't published ML papers or worked at a lab. That is why most of the budget pays for an outside tester and an independent review instead of asking you to take my word for the security claims.
The biggest risk is that the bypasses that matter most are the ones Aegis can't see, such as an agent finding another route to a tool. If that turns out to be most of what the red-team finds, the honest conclusion is that a per-call checkpoint is only useful when paired with sandboxing that removes the other routes. I'd publish that clearly, because it would still save other builders from trusting a checkpoint on its own.
Second, I'm one person, and four months could turn into six. If it slips, I'll say so here rather than go quiet.
Third, the human study may be too small to settle anything. I'll report the uncertainty and won't oversell it.
Fourth, nobody adopts the adapter. Even then, the money leaves behind an open-source checkpoint, a reusable test environment, and a public dataset of real agent bypass attempts that anyone can run their own controls against.
$0 in cash. I've had non-cash support from startup programs: the Mixpanel startup plan, credits from Amplitude, Sentry and Help Scout, and an ElevenLabs startup grant in credits, approved on October 7, 2026. None of it covers model API costs, testers or audits, which is what this grant pays for.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!