01 / OVERVIEW
What it is.
Shovel studies whether an agent operating from a virtual machine can reproduce known security mistakes in a purpose-built practice corpus. It preserves failed attempts, independent reviews and corrections alongside successful runs.
02 / PROJECT TIMELINE
From idea to evidence.
Recorded milestones show how this project developed. Dates describe documented events, not necessarily when the project began.
25 October 2026
Experiment planned
Shovel was scoped as a Greenhouse sibling, using AgenticVM against practice hosts, with Linux first.
26 October 2026
Experiments reviewed
The recorded Linux experiments, controls and independent reviews preserved both successes and a failed attempt.
LATEST RECORDED
Current stage
Research preview. The original guest image is recorded as unavailable for replay, and a Windows target exploit was not completed.
03 / EVIDENCE
What the records show.
The findings index, run transcripts and review-defect ledger record the experiments and their controls. Two campaigns received a conditional measurement verdict.
04 / LIMITATIONS
What remains open.
The host weaknesses were declared in advance, so these are not third-party vulnerability disclosures. The original guest image is recorded as unavailable for replay, and no Windows target exploit was completed.