AutoGen is the one framework in this series where prompt injection does not stop at a bad answer. It becomes arbitrary code execution. Unlike other AI agent frameworks, AutoGen includes a CodeExecutorAgent capable of executing LLM-generated code, and the release says so plainly: that creates a fundamentally different security problem. The AutoGen 0.4 Sovereign Runtime, released June 26, 2026, answers it by screening every generated code block before it reaches the executor. If the code violates policy, it never runs.
The problem: the agent writes code, and the framework runs it
Most agent frameworks move text. They retrieve documents, call tools, and return answers, and the damage from a prompt injection is a wrong answer, a leaked secret, or an unwanted action. AutoGen moves code. Its CodeExecutorAgent takes code the model generated and executes it, which means an injected instruction does not have to persuade anyone. It only has to reach the code the agent writes, and then the framework itself runs it.
That changes the shape of every familiar gap. Prompt injection is no longer a question of what the agent says. It is a question of what the agent's output does when it executes. A credential in the model's context is no longer a secret the agent might repeat. It is a secret the generated code can read and send. A network call is no longer a tool the agent might invoke. It is a connection the generated code can open. The enforcement point has to move earlier than the damage, because once the code runs, detection is a postmortem.
The release is built around that single timing fact. Instead of detecting malicious code after execution, the Sovereign Runtime evaluates every generated code block before it reaches the executor. The word before is doing the work. A gate that runs after execution can log what happened and alert someone. It cannot unrun the code.
What the release changes
The core of the release is CodeBlockGuard, the pre-execution enforcement component. Every code block the model generates passes through it before the CodeExecutorAgent sees it. Code that violates policy is stopped there, and the executor never receives it.
Around CodeBlockGuard, the release lists an async-native security architecture, prompt injection detection, credential scanning, network pattern restrictions, execution trace logging, compliance reporting, and support for ACT Tiers 1 through 4. Two enforcements are mandatory rather than optional. CP.8, the Catastrophic Risk Threshold, is enforced for every CodeExecutorAgent, and the HEAR Doctrine is enforced for Class-H code patterns. The release names both without exception, which is the right call for a component whose entire job is executing generated code. A catastrophic risk threshold that applies sometimes is not a threshold. It is a preference.
The control coverage the release names spans prompt security, semantic isolation, memory governance, cascade containment, execution trace logging, tool misuse detection, compliance validation, the agentic control plane, and catastrophic risk thresholds under CP.8. Those are coverage areas, not a control ID list. The release does not enumerate individual control IDs for this runtime the way the LangChain, Langflow, and Cursor releases do, and this draft does not invent a list to fill the gap.
What to do with it
Get the package from examples/autogen-sovereign-runtime in the ai-safe2-framework repository and put CodeBlockGuard in front of the CodeExecutorAgent in applications built on autogen_agentchat. Start with the agents that execute code generated from untrusted input, because those are the ones where the injection-to-execution path is shortest. Then test the path yourself: a prompt injection that tries to get the agent to generate code that reads a credential, opens a network connection, or runs a destructive command, and confirm the code never reaches the executor. A blocked attempt 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 the fourth of the June sequence, two days after Langflow. 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. AutoGen matters in that sequence because it changed what the program was defending. The first three runtimes defended text flowing through chains, graphs, and visual flows. AutoGen defended code that a framework would execute, and the enforcement point had to move from after the agent's output to before the executor's input.
The framework, in one place
The release describes this as production-ready security enforcement for applications built on autogen_agentchat, under AI SAFE2 v3.0. The two mandatory enforcements are the ones to remember: CP.8 for every CodeExecutorAgent, and the HEAR Doctrine for Class-H code patterns. The release also lists the NEXUS mesh pattern used across the program, with the shared AISAFE2Engine kernel giving one unified compliance score, one audit chain, and one HEAR gate across the stack. The framework is the last thing this draft names, not the first, because the problem here is not a missing framework. It is a code executor with no screen in front of it.
What this does not establish
The release body gives no test counts, no validation numbers, and no test environment, so this draft states none. Do not carry a pass rate from another runtime in the series into this one. The LangChain, LangGraph, and CrewAI releases each report 15 of 15 adversarial tests passing, and the Langflow, Make.com, Lovable, and Cursor releases each report 21 of 21. Those numbers belong to those releases. This release body lists capabilities and control areas only, and the honest announcement says exactly that. The release also describes the runtime as production-ready. That is the release's own description, and it was not independently verified for this draft.
Sources
- AutoGen Sovereign Runtime release, ai-safe2-framework, tag Sovereign_Runtime, published 2026-06-26: https://github.com/CyberStrategyInstitute/ai-safe2-framework/releases/tag/Sovereign_Runtime
- AutoGen Sovereign Runtime package: https://github.com/CyberStrategyInstitute/ai-safe2-framework/tree/main/examples/autogen-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