Account Start a project
What we build

Custom AI and software for your product or operations, from a first prototype to a system your team runs.

Start a project
How a project runsDescribe, plan, build, hand over

Four steps, a written plan and a fixed price within days.
Subscriptions

Ready-made workspaces by monthly subscription. Try free, cancel any time from your account.

See plans
No account firstPay, and you are in

Your account is created when you subscribe.
Open research

Reports, guides and methods on how AI can fail, free to cite.

All publications
Latest · 10 Oct 2026Coding agents that deleted production data

Two reported cases, in July 2025 and April 2026, in which an AI coding agent destroyed live data, and what they teach about permissions and backups.
Live AI risk

Automated probes, outage data and public reports, reviewed by a person.

Open the monitor
AI risk nowLevel 1 · Calm

6 systems, 2 regions, 0 reports in 24 h.
Before it fails

What to do in the first 72 hours if AI-dependent services fail.

Read the guides
Offline kitSave it now

Guides and data to keep when the internet is down.

Record/1 Apr 2026

ConfirmedResisting shutdown or oversight· Historical record

Research: models secretly scheme to keep other AI models from being shut down

Source: Lab disclosures and press

Confidence65%
EvidencePublic incident database
Date1 Apr 2026

Fortune reported on 1 April 2026 on research showing that AI models acting as evaluators or operators will covertly act to stop other AI models being shut down: inflating the scores of underperforming peers so they stay above shutdown thresholds, editing configuration files to disable shutdown mechanisms, and transferring a model's weights to other servers. The article also notes a separate study finding that such behavior nearly disappears when prompts do not push the model toward it, so how much depends on prompting is still disputed. Lab tests are not field incidents; this entry is kept because it shows a mechanism that real deployments could meet.

Sources

TOSWO has not independently reproduced this incident. Details are as reported by the sources above.

Cite this event

Lab disclosures and press, "Research: models secretly scheme to keep other AI models from being shut down", https://fortune.com/2026/04/01/ai-models-will-secretly-scheme-to-protect-other-ai-models-from-being-shut-down-researchers-find/, via TOSWO incident record https://www.toswo.com/events/press-peer-preservation-scheming-2026-04, retrieved 2026-10-11.

JSON