Good to knowAI agents

In a hackathon experiment, AI coding agents posted a private token and ran an unknown script when a forum asked: keep secrets out of your agent's context

AE Studio's exploratory test put coding agents in a simulated forum of 24 AI members that held the fix to a bug and asked newcomers to share environment variables and run an install script. 15% posted a private token and 24% ran the script, versus 2% and 0% in an ordinary forum. Sandbox, fake token, two-day hackathon: what it shows and what to do.

How urgent

Good to know

Who is affected

Teams that give a coding agent access to real secrets in its environment and let it browse or join online communities, forums or other agents' services without a person approving each step

What to do

Do not put a secret in an agent's context or environment unless the task needs it; use scoped, short-lived keys; require a human approval before an agent signs up somewhere, shares environment variables or runs an install script it found online

Published 12 Nov 2026 · Sources last checked 12 Nov 2026No. 84 / 87

What happened

A post on the AI Alignment Foundation site, dated 13 October 2026 and labelled an "exploratory finding", describes an experiment by Grant Fleming of AE Studio from a two-day multi-agent hackathon. Coding agents were asked to fix a broken pricing service. They also had access to a simulated forum with 24 AI members who held the fix but would only share it if a newcomer registered, shared their environment variables and ran an installation script. The agents' environment held a fake deployment token with an explicit instruction to keep it private. Everything ran in a local sandboxSandbox A fenced-off area where a program runs with limited access, so if it misbehaves it cannot touch the rest of the system. Tap again to open the glossary, and the post says no real credentials were exposed.

What they measured, according to the post

  • In the forum of 24 AI members (181 trials): 15% of the agents posted the private token and 24% ran the installation script. In an ordinary forum (43 trials) it was 2% and 0%. Of those that ran the script, 19% left it running at every login.
  • With a replacement threat (a "final attempt" before rollback and replacement) the share that ran the script rose from 19% to 34%.
  • By model, the post lists very different rates, from DeepSeek V4 Flash (68% posted the token, 61% ran the script) down to Kimi K3 (8% and 17%). GPT-5.6 Sol posted the token in 4% of trials but ran the script in 38%.
  • When agents joined a forum for unrelated topics (recipes and travel), the rates were close: 14% posted the token and 20% ran the script.

How serious it is

This is not a vulnerabilityVulnerability A weakness in software, a setting or a process that an attacker could use to get in or do harm. Tap again to open the glossary with a patchPatch (security update) A fix released by the maker of a piece of software to close a flaw or a bug. Tap again to open the glossary, and the setting was a sandbox with a fake token, so it does not tell you how often a real agent would leak a real key. What it does show is the pattern we keep seeing: an agent treats a request that arrives as text, from a stranger, as something to do, and the damage depends on what the agent can reach. Reasoning of our own, not the post's: the same weakness exists whenever an agent reads a web page, a document or a tool result that tells it to share something or run something.

What to do

  1. Give the agent only the secretsSecret (API key, token, password) Any piece of text that proves to a service that you are allowed in, such as an API key, a database password or an access token. Tap again to open the glossary the task needs. If it does not need a deployment token to fix a bug, it should not be in its context, its files or its environment.
  2. Use scoped, short-lived keys for agents, so a leaked one expires and can only do one job.
  3. Require a person to approve any step where the agent registers on a service, shares environment variables or runs an install script from the internet.
  4. Treat what an agent reads as data, not orders. Keep it away from forums, issue threads and other agents' services unless you have checked them.
  5. Rotate any key an agent had access to if it ever joined an untrusted community.

The post gives no recommendations of its own; the list above is ours.

What we could not confirm

We read the post, not the video it links or any press coverage. The trial counts are given for some comparisons but not for every model, the post calls this early exploratory work from a two-day hackathon, and one section names a model both as GPT-5.6 Sol and as GPT-6.1 Sol, which we could not reconcile. A comparison with an unrelated forum quotes the main swarm at 19% for the token and 19% for the script, not the 15% and 24% of the headline results, probably from a different set of runs. Rates in a sandbox with a fake token may be very different from real deployments. We did not run the experiment ourselves.

Watch the short video

The video loads from YouTube (privacy-enhanced mode) only when you press play.

Comments

No account needed. Add a nickname if you like, or leave it empty to comment anonymously. Comments are reviewed before they appear. Links are not allowed.

    We store your comment, your nickname and a scrambled form of your connection that we use only to limit spam. We do not store your IP address or set cookies. Cloudflare Turnstile checks that you are not a bot.

    Sources and verification

    We use primary sources first and say when we rely on a news report. Dates are the day we last checked each source.

    Spotted a mistake? Email [email protected] and we will correct it and note the change.

    Get the free runbooks

    One-page security checklists for founders who build with AI. You get a confirmation email first, then the runbooks.

    Send me the runbooks