Four Additions and One Revision

A second catalog update on October 18 brings our resource catalog to 217 entries. It adds four papers and revises one existing record; nothing is removed.

  • Auditable Claims about AI Agents: argues that a claim about an agent can be checked only when its policy, scope, decision rule, and obtainable records with their writers are fixed before any verdict, then applies that check to six common claims. Preprint. Paper.
  • agent-risk: assesses a proposed change to a deployed agent from its existing logs, replaying logged episodes and rerunning from saved states only those the log cannot settle. Preprint. Paper, code.
  • The Autonomy Tax: compares models trained to resist prompt injection with undefended baselines on 97 agent tasks, and finds that the defended models fail benign tool calls while straightforward attacks still get through. EMNLP 2026. Paper.
  • WeClawArena: a sandbox and benchmark for agents that act for different users over personal workspaces. It audits attack success from recorded runtime evidence, separately from task utility. EMNLP Findings 2026. Paper, code, data.

The CatchBench record now links its paper (arXiv:2608.22808) and is also listed as a paper.

Yue Zhao, who maintains Audit Commons and its companion list, co-authored all five works, so each carries the Maintainer project tag. They entered the companion list under its usual inclusion rules after an independent review. This date records catalog publication; original release dates remain in the linked sources. Updates appear in our RSS / Atom feed.