RESEARCH · RESEARCH · #3801
Reversibility Classes framework and registry for AI agent tool actions (arXiv:2611.22075v1)
The paper defines 'Reversibility Classes', a five-level vocabulary with an explicit prior-state capture contract, and publishes a pinned implementation plus an evidence-linked registry covering 536 tools across 20 MCP servers. Evaluations (738 rule-case checks and 2,158 recorded runs) show the implemented session guard blocked all 29 attributable attack-positive runs while adding nine approval points across 1,209 benign runs; controlled file, Git, SQLite and browser experiments show recovery depends on captured state and route and reveal a Git merge-state exception. The authors say the evidence supports making recovery assumptions explicit and testable but does not establish registry-wide accuracy, reduced production harm, or general recovery reliability.
KEY POINTS
- The paper defines 'Reversibility Classes', a five-level vocabulary with an explicit prior-state capture contract, and publishes a pinned implementation plus an evidence-linked registry covering 536 tools across 20 MCP servers.
- Evaluations (738 rule-case checks and 2,158 recorded runs) show the implemented session guard blocked all 29 attributable attack-positive runs while adding nine approval points across 1,209 benign runs; controlled file, Git, SQLite and browser experiments show recovery depends on captured state and route and reveal a Git merge-state exception.
- The authors say the evidence supports making recovery assumptions explicit and testable but does not establish registry-wide accuracy, reduced production harm, or general recovery reliability.
WHY IT MATTERS
Provides a testable taxonomy, implementation and registry for whether AI-agent tool actions can be undone, giving concrete tooling and empirical results relevant to agent safety and governance while explicitly noting limitations.