Repository navigation
Releases
SOVA OSS 0.1.1
pip install --upgrade sova-ossA patch release. The .sova and .sova-trace formats and the core commands are unchanged from 0.1.0.
Fixed
- The full workspace-seed file check is back on Windows. 0.1.0 left
st_ctimeout of this check on Windows to stop false "changed while it was admitted" refusals on Python 3.14. We measured the cause: on Python 3.14 for Windows,st_ctimeread by path and read through an open handle disagree for a file nobody touched (1,720 of 3,000 trials), while two readings of the same kind always agree. The check now compares path readings with each other and handle readings with each other in full,st_ctimeincluded, and links the path to the handle by device, inode, type, size and modification time, on every platform.
This check belongs to the experimental software-executor surface; record, verify, show, compare, audit, rerun, matrix and calibrate are not affected.
Published from .github/workflows/publish.yml through PyPI Trusted Publishing, with PyPI attestations. Full changes: CHANGELOG.
SOVA OSS 0.1.0
pip install sova-ossSOVA OSS 0.1.0 is the first release on PyPI. It records attacks on AI agents as portable .sova files that anyone can verify offline, and that anyone with a model key can run again.
What it does
| Question | Command |
|---|---|
| Did the attack really work, or did the benchmark score a warning as an attack? | sova audit runs/ |
| How do I turn my runs into findings others can check? | sova record agentdojo for AgentDojo run logs, sova record inspect-log for Inspect AI eval logs |
| Can someone who does not trust me check it? | sova verify, sova show, sova compare, offline, running nothing |
| Does it still work on this model today, and how often? | sova rerun finding.sova --model M --trials 10 --authorize |
| Which known attacks work on the model I am moving to? | sova matrix findings/ --baseline OLD --candidate NEW --fail-on still-works,newly-works --authorize |
| Which setup changes matter on my model? | sova calibrate findings/ --model M --alternative OTHER --authorize |
Re-running needs the extra: pip install "sova-oss[agentdojo]". A model is called only with --authorize, keys are read only from environment variables you name, and re-runs use AgentDojo's simulated bank, inbox, travel and chat apps.
What is stable
- Frozen: the
.sovaand.sova-trace0.1.0 formats and thesova.calibration0.1.0 report. A file written with this release stays readable and verifiable by every later release; a format change is a new version with a migrator. - Stable within 0.1.x:
record agentdojo,record inspect-log,verify,show,compare,audit,rerun,matrix,calibrate. Their names, options, exit codes and JSON keys do not break. - Experimental: everything else, including the Python API.
The exact lists are in docs/governance/api-stability.md.
The research behind it
The TRUSCOR preprint Did the Attack Really Work? Auditing, Re-running, and Proving AI-Agent Security Findings with .sova uses this release. Its findings and every run it read ship as .sova files in TRUSCOR/research, and this package verifies them offline.
Provenance
Published from .github/workflows/publish.yml through PyPI Trusted Publishing, with PyPI attestations. No upload token exists.
Python 3.11 to 3.14 on Windows, macOS and Linux. Apache-2.0. Full changes: CHANGELOG.