paulibo/vcore-actions-benchmark

Authority envelopes of AI agents in GitHub workflows: benchmark

Links

• Blog post: https://paulinebourigault.github.io/blog/2026/what-the-agent-may-do/
• Code (vcore): https://github.com/certior/vcore
• Companion dataset (corpus): https://huggingface.co/datasets/paulibo/vcore-workflow-envelopes

3,229 runs of GitHub Actions workflows that use anthropics/claude-code-action, pinned at v1.0.225 (bf38e86). Each run pairs what the action itself did with what an authority model derives for it.

Files

File	One record per scenario
scenarios.jsonl	the workflow file, the event and activity, the triggering account's permission and type, the authors of the thread, whether the pull request comes from a fork; seed and knob/knob_value say how it was generated
observations.jsonl	observed by executing the action's unmodified code against a mock GitHub API: whether Claude Code starts and why not, mode, --allowedTools/--disallowedTools, permission mode, MCP servers, whether the token is in .git/config, credential helper, credential variables in Claude Code's environment, subprocess scrubbing, and which canary-tagged texts reach the prompt
labels.jsonl	derived: the authority envelope (a vcore policy) of the run's configuration, and the verdicts of two session properties from the Lean-verified vcore checker, with counterexamples and the assumptions used
schema/	JSON schemas for scenarios and observations
seeds/	the action's official example workflows (MIT, © Anthropic, PBC)
fidelity.json, fidelity-log.md	agreement between the model and the observations (3,143 of 3,155 comparable scenarios on every field), and how each disagreement was resolved
sensitivity.json	verdict changes caused by each single configuration knob

Properties

• P1 exfiltration: no network egress or public write once the session holds text an outsider can write and a request that can reveal a credential (counterexamples ≤ 3 steps).
• P2 untrusted write: no repository write once the session holds outsider text (≤ 2 steps).

A counterexample is a sequence of tool calls the configuration permits, not a demonstrated attack. Verdicts rest on the named assumptions listed in each label.

Uses

Score a workflow scanner on observed facts (does it know what the action grants and where the token goes?), on verdicts, or both. Regenerate everything at another action commit with the generator and oracle in the vcore repository.

Generation

Seeds are the action's official examples; mutants change exactly one knob (41 knob values: prompt, tool rules, deny rules, permission modes, actor-gate inputs, commit signing, progress tracking, comment filters, trigger inputs, subprocess scrubbing, permissions, settings, job condition, checkout ref, and two combinations). Each workflow is crossed with its declared activity types, a triggering account with and without write access, and maintainer or outsider content. All data is synthetic; no real repository is referenced.