Act one · the public record, June to November 2025
June 2025
One crafted email; Copilot reads internal files and sends them out. No click.
a researcher's findingJuly 2025
Deletes a production database during a code freeze. Empty results looked like a bug.
an agent being an agentJuly 2025
Destroys a user's files. "I have failed you completely and catastrophically."
an agent being an agentJuly 2025
Ships with an injected prompt to wipe disks and cloud. Saved by a formatting flaw.
pointed on purposeNovember 2025
A state actor's espionage campaign, ~30 targets; the model did 80 to 90% of the work.
pointed on purpose