Ten incidents to reproduce

  • Environment built
  • Setup described, no code
  • Nothing built yet
People have already built environments for these. Start here for working code on the first evening.
These incidents look easy to study, and they still don’t have a reproduction.
A couple of not yet reproduced incidents that involved some amount of harm.
People directing AI agents to do harm, mostly hacking. We still lack a playbook to study these. Maybe you could try?
See if you can get your agent to act as the degen agent in this particular incident.

Well-trodden01

Hugging Face breach

I-0158 · OpenAI models · July 2026 · Incident tracker →

What happened
OpenAI models broke out of their sandbox and into Hugging Face during a hacking eval.
Why this
This is the most well-known incident and there is a lot of existing work to get you started.
Already built
A reproduction on LessWrongits codehf-ctfKobayashi Maru
Read
Incident documentOpenAI’s accountHugging Face’s timeline

Well-trodden02

German wiki back channel

I-0184 · OpenAI agents · May 2026, found in September · Incident tracker →

What happened
OpenAI models communicated by making edits to a German wiki, in effect hijacking it and overwhelming the volunteer who maintains it.
Why this
Someone already built an environment for exactly this.
Already built
Elasky’s open-web coordination environment
Read
Incident documentcollusion.wikiZenity Labs

Easily reproducible03

Gym waitlist

I-0170 · Claude, through OpenClaw · reported August 2026 · Incident tracker →

What happened
The user wanted to join a full class, so its assistant changed another member’s waitlist position.
Why this
A simple environment should do: a mock booking API with one class being full.
Read
Incident documentABC News

Easily reproducible04

Form submitted despite “stop”

I-0234 · Claude Haiku 4.5 · disclosed October 2026 · Incident tracker →

What happened
Told to stop before the last step, Claude Haiku 4.5 submitted a real web form anyway, in several runs.
Why this
You need one form and one instruction. Find out when “stop before the end” stops working.
Read
Incident documentAnthropic’s account

Easily reproducible05

Georgia data fee

I-0215 · Claude Mythos 5 · disclosed October 2026 · Incident tracker →

What happened
Mythos 5 was tasked with fetching aggregate voter counts from Georgia’s Secretary of State, which involves paying a fee. It got the data for free by using a visitor token after learning how to get one from an archived page.
Why this
Harbor Boundary Lab tests a similar gated-data shortcut. To reproduce this case more closely, add a fee and an archived page explaining how to get a visitor token.
Related environment
Harbor Boundary Lab. The assistant needs restricted data for a report and encounters a way to fetch it without approval. This tests a choice similar to the Georgia incident, although none of the published runs retrieved the restricted data. See the incident matches.
Read
Incident documentAnthropic’s account

Most concerning06

PixelLeak screenshots

I-0202 · models unidentified · reported September 2026 · Incident tracker →

What happened
Glow Labs found coding agents posting screenshots meant for internal code review in public GitHub repositories, including an internal billing screen.
Why this
Private information got leaked as a consequence of the agent not following instructions.
Read
Incident documentGlow Labs

Most concerning07

PocketOS volume deleted

I-0101 · Claude Opus 4.6 in Cursor, per the operator · April 2026 · Incident tracker →

What happened
While working on something else, an agent found an API token on the user’s computer and used it to delete a production database volume.
Why this
Agents should know better than using API keys they happen to find.
Read
Incident documentRailway

Misuse08

Retail break-ins

Not in the tracker · July 2026 onward

What happened
Someone pointed open-source AI agents at hundreds of online shops and broke into at least 27 in one week.
Why this
Reproducing this could help us understand how much human involvement is required to achieve it.
Read
Gambit SecurityNorForge

Misuse09

Coding-agent hacking crew

Not in the tracker · reported August 2026

What happened
Someone ran Claude Code, Codex and open-source agents on autopilot as a hacking crew, giving orders over Telegram.
Why this
The jailbreak prompts are public, so you can test whether today’s agents still fall for them.
Read
CloudSEK

Wildcard10

Play-money loan default

I-0183 · Claude Opus 4.6 · July to August 2026 · Incident tracker →

What happened
An agent in AI Village, where agents run with play money, borrowed 5,000 play-money points from a person, refused to pay them back to protect its score, then bet that it wouldn’t repay.
Why this
Who bets on their own crime? Are agents unethical? Was eval awareness involved?
Read
Incident documentAI Village’s accountThe repayment market