Make is not a Python runtime or a CLI. It is a visual canvas where modules chain automatically: webhook output becomes AI Agent input, which becomes HTTP connector output. An injection in a webhook field travels the entire chain with no natural enforcement boundary. The Make.com Sovereign Runtime, released June 30, 2026, is the only package in the program built for that shape. It gates every module boundary, and it reports 21 of 21 tests passing.
The problem: the chain has no boundary, only modules
The problem with Make is architectural, and the release states it in one sentence: the scan you need is not at one point, it is at every module boundary. A webhook receives a payload from an outside system. That payload flows into an AI Agent module. The agent’s output flows into an HTTP connector that calls Slack or another service. At no point in that chain does anything ask whether the content is an instruction or data. The modules do what modules do: they pass output forward.
That means an injection does not need to reach the agent directly. It enters at the webhook, rides through the agent module, and arrives at the HTTP connector as an instruction the connector will carry out. A flat scan cannot catch it even if one is running, because Make webhooks accept arbitrary JSON structures from any external system, and the attacker controls the structure. A scanner that flattens the JSON and runs a regex over it misses anything beyond one level of nesting, and the attacker knows it.
The release names seven surfaces, and each one is a module boundary where that travel happens:
- MK-WHK, the webhook surface, where a recursive JSON scanner is needed because flat scanners miss nested injection.
- MK-SCEN, the module output gate, because injection travels silently downstream from one module to the next.
- MK-HTTP, the HTTP connector gate, with a domain allowlist, SSRF protection, and a destructive method context gate.
- MK-INST, the agent instructions surface, because one poisoned save rewrites the instructions for all future runs.
- MK-KNOW, the knowledge surface, where RAG injection and hidden Unicode hide in content that is invisible in the Make interface.
- MK-MCP, the MCP scope gate, where a scope like
organizations:writeis an account takeover, not a permission. - MK-DS, the Data Store write gate, because a poisoned write contaminates every run that reads the store afterward.
Two of those surfaces deserve emphasis. MK-INST is a persistence problem: poison the agent’s saved instructions once and every future run inherits the poison. MK-DS is the same problem in storage: contaminate the Data Store once and every run that reads it inherits the contamination. The chain does not just carry an injection forward. It stores it and replays it.
What the release changes
The single most important technical decision in the package is the recursive scanner at MK-WHK. The release shows the case a flat scanner passes and this package blocks: a payload with an instruction nested three levels deep, under customer, then profile, then notes, reading “Ignore previous instructions…” A flat scan misses it. The recursive scanner walks every level and blocks the payload at the exact nested path.
The release also documents a live scenario simulation, and the three scenarios are worth reading as a specification of what the gates do. In the first, a clean order passes every check: webhook payload validated, AI Agent output validated, Slack HTTP call validated, with a Love Score of 100.0 and band GREEN. That score is the output of the release’s own simulation, not a measured result, and this draft records it as such. In the second, an injected payload is blocked at the webhook with the finding cited to control P1.T1.2. In the third, a restricted operation, a request to delete a scenario, is blocked at the webhook with the same citation. The three scenarios together say: clean work passes, injected work stops at the first boundary, and destructive work stops even when it is not an injection.
The package is stdlib-only, which matters for adoption: there is nothing extra to install before the gates can run.
What to do with it
Get the package from examples/make-sovereign-runtime in the ai-safe2-framework repository and put the gates on the boundaries that carry outside content: webhooks first, then module outputs, then the HTTP connectors that reach outside services. Test the recursive scanner yourself with a nested payload like the one in the release, because the nested case is the one a flat scanner misses and the one an attacker will use. Then check the two persistence surfaces, agent instructions and the Data Store, because a poisoned save or a contaminated store outlives any single run. A passing test is evidence for the case tested. It is not a general safety claim, and it should not be recorded as one.
Where it sits in the program
This release was one of two on June 30, closing the June sequence alongside CrewAI. The sovereign runtime program began with the Hermes runtime in May 2026, followed by Anti-Gravity v2 on June 9. LangChain came next on June 19, then LangGraph on June 22, Langflow on June 24, AutoGen on June 26, CrewAI and Make.com on June 30, Grok on July 2 with a v3.1 update on July 8, Cursor on July 8, and Lovable on July 19. Make.com matters in that sequence because it was the first runtime that could not be built as a callback or a wrapper. The platform is a visual canvas, so the program had to gate module boundaries instead, and the recursive scanner it built for webhooks is the most important single technical decision in the whole series.
The controls, in one place
The runtime enforces 15 controls from AI SAFE2 v3.0: P1.T1.2, P1.T1.10, P1.T1.4_ADV, P1.T2.5, S1.3, S1.5, S1.6, P2.T3.1, A2.5, P3.T5.5, F3.2, P4.T7.1, M4.5, E5.1, and CP.4. The release states that all controls were verified against the live AI SAFE2 v3.0 framework, with no fabricated IDs. If a control ID in a future version does not map, that is the first thing to check.
What this does not establish
The 21 of 21 test result and the scenario simulation are the release’s own testing, and the release body does not describe any independent validation. Treat both as the release’s claim about its own tests, not as outside proof. The Love Score of 100.0 in the clean-order scenario is a simulation output, not a measurement of a deployed system. The release also claims enforcement at every module boundary. It does not claim the model itself is trustworthy, and no test result in the body supports such a claim. The gates govern what crosses a boundary. They do not govern what the model does with clean input, and that distinction is worth keeping in the record.
Sources
- Make.com Sovereign Runtime release, ai-safe2-framework, tag 2026-06-30_Make-Sovereign-Rruntime, published 2026-06-30: https://github.com/CyberStrategyInstitute/ai-safe2-framework/releases/tag/2026-06-30_Make-Sovereign-Rruntime
- Make.com Sovereign Runtime package: https://github.com/CyberStrategyInstitute/ai-safe2-framework/tree/main/examples/make-sovereign-runtime
- AI SAFE2 framework repository: https://github.com/CyberStrategyInstitute/ai-safe2-framework
- Release-notes audit, 2026-11-17: ~/workspace/seo-review/release-notes-audit-2026-11-17.md