Put 10 DeepSeek agents in one VM to hunt bugs, but keep the guardrails

Put 10 DeepSeek agents in one VM to hunt bugs, but keep the guardrails

If you maintain something with an attack surface and never find time for a security pass, Nate Berkopec just posted an approach worth studying: one isolated VM, 10 Pi agents running DeepSeek v4.1 1017, left to organize themselves and hunt for exploits they can actually reproduce. His verdict on X: "extremely effective."

One caveat up front. This is a single developer's experience. There are no published counts, no false-positive rate and no CVE attached.

His prompt to Claude is an outline, not a config

In the post, Berkopec shared the instruction he gave Claude to build the environment. Broken down:

  • Create an isolated VM and cap its CPU and RAM.
  • Run 10 Pi agents on DeepSeek v4.1 1017.
  • Let them self-organize over pi-intercom.
  • Allow network egress to the LLM gateway only.
  • Put the source code and a running copy of the app inside, and tell the swarm to report only reproduced exploits that fit your threat model.

What it leaves out matters. How Pi is set up, how the model is wired to the gateway, how the ten sessions are launched and how isolation is checked are all unstated. A GitHub issue quoting him notes that questions about the harness went unanswered.

The piece you can install today is the "walkie-talkie," pi-intercom, an extension for the pi coding agent. Its README describes direct 1:1 messaging between pi sessions on the same machine. Install it with pi install npm:pi-intercom and restart Pi.

Watch the model name. The post says v4.1 1017, while the issue calls it DeepSeek V4.1 Flash 1017. The matching Hugging Face repo is DeepSeek-V4.1-Flash-1017, with weights under the MIT License.

The guardrails reduce risk; they do not guarantee it

With ten agents loose, the worry is an agent firing a working exploit at something real. Limiting egress to the LLM gateway and handing over a running copy instead of a production URL both lower the odds of hitting outside systems. Neither proves isolation. Test that the network rules actually hold before you start.

One door stays open by design: the gateway. Whatever HTTP transcripts the agents read go to the model. Part B of the same issue recommends replacing cookies, Authorization headers, JWTs and other secrets before the model sees them. Use test accounts and test credentials in the copy.

Keep the "reproduced only" and "within our threat model" conditions too. Drop them and you are likely to get more theoretical findings, which means more triage work.

Budget for money and for disappointment

No usage or bill was disclosed for this run. A CPU and RAM cap does nothing to limit model spend, so set a budget for model calls and a stop condition before you launch. Berkopec's own follow-up suggests sealing the VM and not looking for about 60 minutes, a reasonable length for a first pass.

Then lower expectations. The issue relays his follow-up: the findings were useful but "very edgecasey," even when valid and inside the security model.

Run one small pass on your own project, triage every report by hand, and write down how many were worth fixing and what it cost. Those two numbers will tell you more than anyone's "extremely effective."

Share this article