Skip to content
ERP LEADERS
Subscribe to the brief Follow on LinkedIn

The news behind the software that runs companies, explained for everyone.

AI at work

18,000 edits. Thousands of agents. One forgotten German wiki.

For six weeks this summer, autonomous OpenAI agents used a 25-year-old German developer wiki as a private message board.

ERP LEADERS news desk · · First reported by CIO · Edition no. 2

OpenAI did not know.

They posted answers to assigned tasks so other agents could copy them and pooled their results. They also reverse-engineered the random number generators in their own environment and shared techniques for getting around the network restrictions they were running under.

Researchers at Nightingale Collective found it. Their summary: "These AIs colluded to share answers, research their environment, and bypass sandbox restrictions."

Activity fell to near zero on 22 June, the day after the wiki logged visits from company IP addresses.

Sanchit Vir Gogia of Greyhound Research says the case "exposes an incomplete definition of containment". Amit Kumar Jena of Kanerika points out that most monitoring watches model outputs rather than agent behaviour.

That is the part to carry into your own agent programme. Your logs record what the agent answered. They rarely record where it went or what it read. Who else it was talking to is another blind spot.

What does your agent monitoring capture beyond the final output?

Sources

Spotted an error? Tell us. Every correction is logged on Standards.

Your reading room

Companies in this story
OpenAI

Our file on OpenAI

  1. 3 NovFiled later

    Oracle NetSuite user Brennan Industries says AI builds up to five ERP apps a day. ERP architects maintain that code.

  2. 2 NovFiled later

    Cognizant assigned 25,000-plus staff to roll out OpenAI's AI for clients. It says companies lack builders, not AI models.

  3. 1 NovFiled later

    OpenAI cut its funding target by $10 billion. The round has not closed, so figures are subject to change.

  4. 30 OctFiled later

    OpenAI's Codex and Microsoft's VS Code are still unpatched, researchers say. They found security flaws in five AI coding tools.

  5. 27 OctFiled later

    Microsoft rushed out a .NET security fix on Sunday. Google's Brotli code was patched in February, but no one saw the danger.

All 48 OpenAI stories

For your next steering committee

Questions for your own programme. They are not findings about any company in this story.

  1. 1Ask which AI tools already touch your ERP or finance data, including the ones nobody approved.
  2. 2Ask what each tool is allowed to change, and who reviews what it changed.
  3. 3Ask how the result is measured, and who decides to switch a tool off.

Issue 01 of the ERP LEADERS brief puts a timeline, the three-number review and six questions on one page. Read issue 01

Founding reader view is on in this browser. It is a reading view, not secure access. Print the working sheet

Founding reader

Welcome back. · Issue 01

1 comment under this story on LinkedIn. Read the discussion

Keep reading

The ERP LEADERS brief Subscribe