syzkaller AI agentic framework and MCP server

[Dmitry Vyukov]

Hi,

FYI if you are interested in agentic research, and or contributing to
syzkaller/syzbot AI efforts.
We now have own extensible agentic framework called aflow:
https://github.com/google/syzkaller/tree/master/pkg/aflow

It's probably easier to grasp by looking at concrete AI agentic
workflows we have.
One allows to generate fix patches for kernel bugs:
https://github.com/google/syzkaller/blob/master/pkg/aflow/flow/patching/patching.go
Other two analyze KCSAN and KASAN bugs:
https://github.com/google/syzkaller/blob/master/pkg/aflow/flow/assessment/kcsan.go
https://github.com/google/syzkaller/blob/master/pkg/aflow/flow/assessment/moderation.go

The framework includes a set of tools for kernel source code browsing/editing:
https://github.com/google/syzkaller/tree/master/pkg/aflow/tool
and we plan to extend the set with debugger, tracing tools, etc.
The framework also includes "actions" to checkout/build kernels, run
reproducers, etc:
https://github.com/google/syzkaller/tree/master/pkg/aflow/action

The framework is now deployed to syzbot:
https://syzbot.org/upstream/ai
and example of KCSAN assesment workflow:
https://syzbot.oarg/ai_job?id=af54e338-4735-44a3-81fd-5f2f3b376208
and fixing patch generation workflow:
https://syzbot.org/ai_job?id=48473ed4-1f09-40df-a12f-da5f303e29f4

The framework tools/actions are also exposed via an MCP server, and
can be used with e.g. gemini-cli:
https://github.com/google/syzkaller/blob/master/docs/mcp.md
The current status of the MCP server is more of a POC. That's not the
main usage mode for now, and it still has some rough edges. But that
can be improved if necessary.